Enhancing Safety and Capacity in an Adaptive

2 downloads 0 Views 9MB Size Report
Bell Road arterial network: Comparison of number of stops. ...... (detector data and phase timing information) and downloading new timing ...... 6300 - 7200. 33.8.
Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

July 2013

Authors:

Ziad A. Sabra, Ph.D., P.E., PTOE, Sabra, Wang & Associates, Inc. Douglas Gettman, Ph.D., Kimley-Horn and Associates, Inc. Venkata Nallamothu, EIT, Sabra, Wang & Associates, Inc. Caroline Pecker, Sabra, Wang & Associates, Inc.

Prepared by: 7055 Samuel Morse Dr., Suite 100 Columbia, MD 21046 www.sabra-wang.com

 

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2 July 2013

Authors: Ziad A. Sabra, Ph.D., P.E., PTOE, Sabra, Wang & Associates, Inc. Douglas Gettman, Ph.D., Kimley-Horn and Associates, Inc. Venkata Nallamothu, EIT, Sabra, Wang & Associates, Inc. Caroline Pecker, Sabra, Wang & Associates, Inc.

Prepared by: 7055 Samuel Morse Dr., Suite 100 Columbia, MD 21046

i

ii

TECHNICAL REPORT INFORMATION Title: Enhancing Safety and Capacity in an Adaptive Signal Control System—Phase 2 Authors: Ziad A. Sabra, Ph.D., P.E., PTOE, Douglas Gettman, Ph.D., Venkata Nallamothu, EIT, and Caroline Pecker Prepared by: Sabra, Wang & Associates, Inc. 7055 Samuel Morse Drive, Suite 100 Columbia, Maryland 21046 FHWA Contracting Officer’s Technical Representative (COTR): Mr. Joe Bared, Ph.D. Abstract: This research, performed under the FHWA Small Business Innovation Research (SBIR) program, comprised two phases: Phase 1 was completed and published in August 2010 under publication No. FHWA-HRT-10-038. Phase 1 examined the relationships between signal timing and surrogate measures of safety, namely the frequency of rear-end, angle and lane-change conflicts. The FHWA Surrogate Safety Assessment Methodology (SSAM) was used to evaluate various simulated scenarios to test the relationships between signal timing parameters such as cycle, offset, split, phase change interval, detector extension time, left-turn phase protection options and left-turn phase sequence and the occurrence of traffic conflicts. This report summarizes the effort performed under Phase 2 and discusses the development and testing and verification of a safety performance function and a multi-objective optimization methodology using four principle algorithms that comprise the proposed adaptive system for tuning the cycle length, splits, offsets, and left-turn phase sequence at signalized intersections. The safety performance function was developed by training a cascade feed forward neural network to learn the relationship between the signal timing settings, efficiency measures, and the resulting average traffic conflict rates. The average traffic conflict rates were post-processed using SSAM from trajectory data from the VISSIM traffic simulation model. The Virtual D4 controller was used to implement the signal operation in VISSIM. From training the network with approximately 150 scenarios, the average error on the prediction of traffic conflicts on cases that were not used for training was 17%. This average error represented a reasonable performance to use as the safety performance prediction function in the adaptive control system. MATLAB was used to implement the neural network and export the network calculations as a C++ DLL that was integrated with the rest of the adaptive system code.

iii

SI* (MODERN METRIC) CONVERSION FACTORS APPROXIMATE CONVERSIONS TO SI UNITS Symbol

When You Know

in ft yd mi

inches feet yards miles

Multiply By LENGTH 25.4 0.305 0.914 1.61

To Find

Symbol

millimeters meters meters kilometers

mm m m km

square millimeters square meters square meters hectares square kilometers

mm 2 m 2 m ha 2 km

AREA 2

in 2 ft 2 yd ac 2 mi

square inches square feet square yard acres square miles

645.2 0.093 0.836 0.405 2.59

fl oz gal ft3 3 yd

fluid ounces gallons cubic feet cubic yards

oz lb T

ounces pounds short tons (2000 lb)

o

Fahrenheit

fc fl

foot-candles foot-Lamberts

lbf 2 lbf/in

poundforce poundforce per square inch

Symbol

When You Know

mm m m km

millimeters meters meters kilometers

2

VOLUME 29.57 milliliters 3.785 liters 0.028 cubic meters 0.765 cubic meters 3 NOTE: volumes greater than 1000 L shall be shown in m

mL L m3 3 m

MASS 28.35 0.454 0.907

grams kilograms megagrams (or "metric ton")

g kg Mg (or "t")

TEMPERATURE (exact degrees) F

5 (F-32)/9 or (F-32)/1.8

Celsius

o

lux candela/m2

lx cd/m2

C

ILLUMINATION 10.76 3.426

FORCE and PRESSURE or STRESS 4.45 6.89

newtons kilopascals

N kPa

APPROXIMATE CONVERSIONS FROM SI UNITS Multiply By LENGTH 0.039 3.28 1.09 0.621

To Find

Symbol

inches feet yards miles

in ft yd mi

square inches square feet square yards acres square miles

in ft2 2 yd ac 2 mi

fluid ounces gallons cubic feet cubic yards

fl oz gal 3 ft 3 yd

ounces pounds short tons (2000 lb)

oz lb T

AREA 2

mm m2 2 m ha 2 km

square millimeters square meters square meters hectares square kilometers

0.0016 10.764 1.195 2.47 0.386

mL L 3 m 3 m

milliliters liters cubic meters cubic meters

g kg Mg (or "t")

grams kilograms megagrams (or "metric ton")

o

Celsius

2

VOLUME 0.034 0.264 35.314 1.307

MASS

C

0.035 2.202 1.103

TEMPERATURE (exact degrees) 1.8C+32

Fahrenheit

o

foot-candles foot-Lamberts

fc fl

F

ILLUMINATION lx 2 cd/m

lux 2 candela/m

N kPa

newtons kilopascals

0.0929 0.2919

FORCE and PRESSURE or STRESS 0.225 0.145

poundforce poundforce per square inch

lbf 2 lbf/in

*SI is the symbol for the International System of Units. Appropriate rounding should be made to comply with Section 4 of ASTM E380. (Revised March 2003)

iv

TABLE OF CONTENTS EXECUTIVE SUMMARY .......................................................................................................... 1 1 INTRODUCTION...................................................................................................................... 5 1.1 SUMMARY OF PHASE I RESEARCH RESULTS...................................................................... 5 1.2 RESEARCH OBJECTIVES ............................................................................................................ 6 1.2.1 Develop the Design of Experiments (DoE) ................................................................................. 6 1.2.2 Develop the multi-objective adaptive algorithms ........................................................................ 6 1.3 ORGANIZATION OF THE REPORT ........................................................................................... 7

2 EXPERIMENTAL DESIGN..................................................................................................... 9 2.1 EXPERIMENTAL DESIGN FOR THE DEVELOPMENT OF A SAFETY PERFORMANCE FUNCTION ............................................................................................................. 9 2.2 Summary.......................................................................................................................................... 25

3 DEVELOPMENT OF SAFETY PERFORMANCE FUNCTION ...................................... 27 3.1 OVERVIEW OF NEURAL NETWORKS ................................................................................... 27 3.2 SUMMARY ..................................................................................................................................... 36

4 DEVELOPMENT OF MULTI-OBJECTIVE ALGORITHMS .......................................... 37 4.1 OVERVIEW .................................................................................................................................... 37 4.2 FIELD DETECTOR REQUIREMENTS ..................................................................................... 38 4.3 ADAPTIVE ALGORITHMS ......................................................................................................... 39 4.3.1 Cycle Tuning .............................................................................................................................. 40 4.3.2 Cycle Selection .......................................................................................................................... 41 4.3.3 Offset Tuning ............................................................................................................................. 43 4.3.4 Split Tuning ............................................................................................................................... 49 4.4 PHASE SEQUENCE CHANGES.................................................................................................. 57 4.5 PROTECTED / PERMITTED LEFT-TURN MODE CHANGES (DEFERRED FOR FUTURE CONSIDERATION) ......................................................................... 59 4.6 Adaptive Control Algorithms summary ....................................................................................... 60

5 SYSTEM PARAMETERS AND SYSTEM USER INTERFACE ....................................... 65 5.1 SYSTEM ARCHITECTURE......................................................................................................... 65 5.2 SYSTEM PARAMETERS ............................................................................................................. 66 5.2.1 Cycle Time Parameters .............................................................................................................. 67 5.2.2 Enable Safety Evaluation ........................................................................................................... 68 5.2.3 Phase Sequence Parameters ....................................................................................................... 68 5.2.4 Protected/Permitted Phase Parameters (currently disabled and subject to future research) ....... 69 5.2.5 Offset Parameters ....................................................................................................................... 69 5.2.6 Split Parameters ......................................................................................................................... 69 5.3 CONTROLLER CONFIGURATION .......................................................................................... 70 5.4 DETECTORS CONFIGURATION .............................................................................................. 71 5.5 LINKS CONFIGURATION .......................................................................................................... 73 5.6 PHASING DATA ............................................................................................................................ 74 5.7 RING SEQUENCES ....................................................................................................................... 75 5.8 TIME-OF-DAY (TOD) PATTERNS ............................................................................................ 76 v

5.9 PATTERNS SCHEDULE .............................................................................................................. 76 5.10 VISSIM SIMULATION PARAMETERS (Simulation-In-The-Loop) ..................................... 77 5.10.1 Model Geometry ...................................................................................................................... 78 5.10.2 Traffic Demand Data ............................................................................................................... 79 5.10.3 Traffic Control Data ................................................................................................................. 79 5.10.4 Simulation Parameters ............................................................................................................. 80 5.11 SYSTEM PERFORMANCE OUTPUT ...................................................................................... 80 5.11.1 Offset Adjustment Outputs ...................................................................................................... 84 5.11.2 Phase Sequence Adjustment outputs........................................................................................ 87 5.11.3 Split Adjustment Outputs ......................................................................................................... 88 5.11.4 Cycle Selection Outputs ........................................................................................................... 91 5.12 Summary........................................................................................................................................ 93

6 Algorithms Verification ........................................................................................................... 95 6.1 Overview .......................................................................................................................................... 95 6.2 simulation METHODOLOGY....................................................................................................... 95 6.3 Test scenarios .................................................................................................................................. 95 6.3.1 Baseline Arterial Model ............................................................................................................. 95 6.3.2 Analysis Scenario 1 (VISSIM-Virtual D4) ................................................................................ 99 6.3.3 Analysis Scenario 2 with VISSIM-Virtual D4-Kadence (SPF disabled) ................................... 99 6.3.4 Analysis Scenario 3 with VISSIM-Virtual D4-Kadence (SPF enabled).................................... 99 6.4 Testing Cases and Results .............................................................................................................. 99 6.4.1 Case 1, Non-optimum offset selection at intersection M ......................................................... 103 6.4.2 Case 2, Non-optimum offset selection at all intersections ....................................................... 104 6.4.3 Case 3, Over-saturated left-turn (EB) at intersection M .......................................................... 106 6.4.4 Case 4, Over-saturated left-turns (EB and WB) at intersection M........................................... 108 6.4.5 Cases 5 through 7, Non-optimum phase sequence selection at intersection M ...................... 111 6.4.6 Modified baseline for cycle length adjustments....................................................................... 112 6.4.7 Case 8 - Higher Demand Starts Late ........................................................................................ 114 6.4.8 Case 9 – High Demand Starts Early......................................................................................... 116 6.4.9 Case 10- Low Demand Starts Late........................................................................................... 118 6.4.10 Case 11- Low Demand Starts Early ....................................................................................... 121 6.5 Sensitivity Testing of System Parameters ................................................................................... 126 6.5.1 Sensitivity Analysis for Offset Adjustment Parameters ........................................................... 127 6.6 Summary........................................................................................................................................ 135

7 SIMULATION OF REAL-LIFE CASE SCENARIOS ...................................................... 137 7.1 ARTERIAL APPLICATION ...................................................................................................... 137 7.1.1 Simulation-in-Loop Model Development ................................................................................ 139 7.1.2 Simulation-in-the-Loop Assumptions ...................................................................................... 142 7.1.3 Performance Measures ............................................................................................................. 143 7.2 GRID APPLICATION ................................................................................................................. 150 7.2.1 Simulation-in-Loop Model Development ................................................................................ 150 7.2.2 Simulation-in-the-Loop Assumptions ...................................................................................... 153 7.2.3 Performance Measures ............................................................................................................. 154 7.3 SUMMARY ................................................................................................................................... 160 vi

8 FUTURE RESEARCH AND RECOMMENDED ENHANCEMENTS ........................... 163 REFERENCES .......................................................................................................................... 165

vii

LIST OF FIGURES Figure 1. Illustration. Test network for experimental design ...................................................................... 10 Figure 2. Illustration. Variable and fixed arrival volumes .......................................................................... 11 Figure 3. Illustration. Right turn routes....................................................................................................... 12 Figure 4. Illustration. Through routes ......................................................................................................... 13 Figure 5. Illustration. Left turn routes ......................................................................................................... 14 Figure 6. Illustration. Side street approach routes....................................................................................... 15 Figure 7. Illustration. Link lengths modified in experimental design ......................................................... 15 Figure 8. Illustration. Splits adjusted in the experimental design ............................................................... 16 Figure 9. Illustration. Offsets adjusted in the experimental design............................................................. 17 Figure 10. Illustration. Basic concept of a “neuron” in a neural network ................................................... 28 Figure 11. Illustration. Sigmoid activation function ................................................................................... 28 Figure 12. Illustration. Illustration of the training process.......................................................................... 29 Figure 13. Illustration. Neural network training tool in MATLAB ............................................................ 30 Figure 14. Illustration. Stop bar and advance detection .............................................................................. 39 Figure 15. Illustration. Flow chart of cycle time tuning algorithm ............................................................. 41 Figure 16. Illustration. Concept of implementing the next cycle time earlier or later than scheduled ....... 42 Figure 17. Illustration. Cycle selection process .......................................................................................... 43 Figure 18. Illustration. Typical flow profile detector locations on coordinated approaches....................... 44 Figure 19. Illustration. Example of traffic volume count and occupancy data from a typical advance detector ................................................................................................................... 45 Figure 20. Illustration. Example of phase timing for each of the last several cycles .................................. 46 Figure 21. Illustration. Example of average cyclic volume and occupancy profiles .................................. 47 Figure 22. Illustration. Offset adjustment algorithm flow chart ................................................................. 48 Figure 23. Illustration. Ring diagram with barriers..................................................................................... 49 Figure 24. Illustration. Detector layout ....................................................................................................... 52 Figure 25. Illustration. Measuring phase utilization for coordinated-actuated controllers ......................... 53

viii

Figure 26. Illustration. Utilization of phases before and after split adjustment .......................................... 55 Figure 27. Illustration. Flow chart of the split optimization process including safety analysis .................. 57 Figure 28. Illustration. Adjustment algorithms flow chart. ......................................................................... 62 Figure 29. Illustration. Kadence Software-in-the-Loop Simulation ........................................................... 65 Figure 30. Illustration. Kadence field deployment schematics ................................................................... 66 Figure 31. Illustration. Kadence System Parameters .................................................................................. 67 Figure 32. Illustration. Controller configuration parameters ...................................................................... 70 Figure 33. Illustration. Detector configuration parameters ......................................................................... 72 Figure 34. Illustration. Link definition parameters ..................................................................................... 74 Figure 35. Illustration. Phase configuration parameters ............................................................................. 75 Figure 36. Illustration. Ring-barrier sequence ............................................................................................ 75 Figure 37. Illustration. Pattern definition parameters ................................................................................. 76 Figure 38. Illustration. TOD plan parameters ............................................................................................. 77 Figure 39. Illustration. SIL Data Flow ........................................................................................................ 78 Figure 40. Illustration. VISSIM Study Network ......................................................................................... 79 Figure 41. Illustration. Typical Kadence log file illustration ...................................................................... 81 Figure 42. Illustration. Typical output graphs............................................................................................. 83 Figure 43. Illustration. Offset adjustment log (with SPF enabled) ............................................................. 84 Figure 44. Illustration. Offset adjustment and safety prediction output graphs .......................................... 86 Figure 45. Illustration. Sequence adjustment log ........................................................................................ 87 Figure 46. Illustration. Sequence adjustment graph .................................................................................... 88 Figure 47. Illustration. Split adjustment log ............................................................................................... 88 Figure 48. Illustration. Split adjustment Ring 1 .......................................................................................... 89 Figure 49. Illustration. Split adjustment Ring 2 .......................................................................................... 90 Figure 50. Illustration. Split surrogate safety measures prediction graph ................................................... 90 Figure 51. Illustration. Log illustration of plan switching from higher to lower cycle ............................... 91 Figure 52 . Illustration. Graphical illustration of plan switching from higher to lower cycle .................... 92 ix

Figure 53. Illustration. Graphical illustration of incremental adjustment from higher to lower cycle........ 92 Figure 54. Illustration. Baseline arterial configuration ............................................................................... 96 Figure 55. Lane configuration and detection location. ............................................................................... 96 Figure 56. Illustration. Demand and signal timing for the baseline scenario.............................................. 97 Figure 57. Illustration. Node locations...................................................................................................... 101 Figure 58. Illustration. Travel time segments. .......................................................................................... 102 Figure 59. Illustration. Demand and signal timing for Case 1. ................................................................. 103 Figure 60. Illustration. Demand and signal timings for Case 2 ................................................................ 105 Figure 61. Illustration. Demand and signal timings for Case 3. ............................................................... 106 Figure 62. Split adjustments for Case 3. ................................................................................................... 107 Figure 63. Illustration. Demand and signal timings for Case 4 ................................................................ 108 Figure 64. Illustration. Split adjustments for Case 4................................................................................. 109 Figure 65. Illustration. Demand and signal timing information for scenario with non-optimum offset at all intersections. ............................................................................................ 113 Figure 66. Illustration. TOD schedule and demand intervals showing an extended period of low traffic demand. ............................................................................................................ 114 Figure 67. Illustration. Kadence graph for cycle change in Case 8 .......................................................... 115 Figure 68. Illustration. TOD schedule and demand intervals showing an abrupt increase in demand earlier than anticipated. ..................................................................................... 116 Figure 69. Illustration. Kadence graph of cycle change in Case 9 ............................................................ 117 Figure 70. Illustration. TOD schedule and demand intervals showing an extended period of high demand. ..................................................................................................................... 119 Figure 71. Illustration. Kadence graph for cycle change in Case 10 ........................................................ 120 Figure 72. Illustration. Demand and signal timing information for scenario with non-optimum offset at all intersections. ............................................................................................ 121 Figure 73. Illustration. Kadence graph for cycle change in Case 11 ........................................................ 122 Figure 74. Illustration. Bell Road arterial network. .................................................................................. 137 Figure 75. Illustration. Critical routes during game overlaid with peak flows ......................................... 138 Figure 76. Illustration. Bell Road arterial network traffic volumes. ......................................................... 139 x

Figure 77. Illustration. VISSIM snap shot of the Bell Road arterial network........................................... 140 Figure 78. Illustration. Bell Road arterial network- Pattern 1 green times. .............................................. 141 Figure 79. Illustration. Kadence system parameters for the Bell Road arterial network. ......................... 142 Figure 80. Illustration. Bell Road/Bullard Avenue-Kadence offset adjustments. ..................................... 145 Figure 81. Illustration. Bell Road/Bullard Avenue- Kadence split adjustments. ...................................... 145 Figure 82. Illustration. Bell Road arterial network- Average delay time comparison. ............................. 146 Figure 83. Illustration. Bell Road arterial network-Total delay time comparison. ................................... 147 Figure 84. Illustration. Bell Road arterial network: Comparison of number of stops. ............................. 148 Figure 85. Illustration. Bell Road arterial network- comparison of total travel time. ............................... 149 Figure 86. Illustration. DC Grid network .................................................................................................. 150 Figure 87. Illustration. DC Grid network- VISSIM model. ...................................................................... 151 Figure 88. Illustration. DC Grid network turning movement volumes. .................................................... 151 Figure 89. Illustration. DC Grid network- Pattern 1 green times .............................................................. 152 Figure 90. Illustration. DC Grid network- Kadence system parameters ................................................... 154 Figure 91. Illustration. 14th Street, E Street and Pennsylvania Avenue NW: Kadence offset adjustments .............................................................................................................. 155 Figure 92 . Illustration. 14th St, E St and Pennsylvania Ave NW- Kadence split adjustments ................ 156 Figure 93. Illustration. Grid network: Average delay comparison ........................................................... 157 Figure 94. Illustration. Grid network- Total delay time comparison. ....................................................... 158 Figure 95. Illustration. Grid network- Comparison of numbers of stops. ................................................ 159 Figure 96. Illustration. Grid Network- Comparison of total travel time. .................................................. 160

xi

LIST OF TABLES Table 1. Input factors and ranges for a three-intersection model: W (west), M (middle), E (east) ........... 18 Table 2. Control parameters of 100 test cases ............................................................................................ 22 Table 3. Left turn treatment ........................................................................................................................ 25 Table 4. Performance of neural network on training of cases ..................................................................... 32 Table 5. Variables used in split utilization .................................................................................................. 54 Table 6. Example utilization of phases before and after split adjustment................................................... 56 Table 7. Rules to evaluate to consider changing phase sequence ............................................................... 58 Table 8. Rules to evaluate when considering changing left-turn treatment ................................................ 60 Table 9. Change and clearance intervals for the baseline model. ............................................................... 97 Table 10. Kadence system parameters used for verification process. ......................................................... 98 Table 11. Signal Controller parameters used for verification process ........................................................ 98 Table 12. Test cases and respective adaptive algorithm ........................................................................... 100 Table 13. Node evaluation data................................................................................................................. 101 Table 14. Network Performance data. ...................................................................................................... 101 Table 15. Filtering thresholds for conflicts in SSAM. .............................................................................. 102 Table 16. Case 1 results - MOE collected at intersection M (through lanes only) ................................... 104 Table 17. Case 2 results - MOE collected mainline arterial ..................................................................... 105 Table 18. Case 3 results – MOE collected at intersection M .................................................................... 107 Table 19. Case 4 results - MOE collected at intersection M (EBL and WBL) ......................................... 110 Table 20. EB Left-turns - Average Delays per Vehicle at intersection M – Safety Disabled ................... 110 Table 21. WB Left-turns - Average Delays per Vehicle at intersection M – Safety Disabled ................. 111 Table 22. Average Delay per Vehicle at intersection M - No Adjustments.............................................. 111 Table 23. Average Delays per Vehicle at Intersection M - Kadence Safety Disabled ............................. 112 Table 24. Case 6 Results - Queue Lengths at intersection M ................................................................... 112 Table 25. TOD schedule for modified baseline case. ............................................................................... 114 Table 26. Demand interval schedule for modified baseline case. ............................................................. 114 xii

Table 27. Case 8 Results – Average Delays per Vehicle at each intersection .......................................... 115 Table 28. Case 9 Results, average delays per vehicle at each intersection ............................................... 117 Table 29. TOD schedule for modified baseline case. ............................................................................... 119 Table 30. Demand interval schedule for modified baseline case. ............................................................. 119 Table 31. Case 9 Results – Average Delays per Vehicle at each intersection .......................................... 120 Table 32. Demand interval schedule for modified baseline case. ............................................................. 121 Table 33. Case 11 Results – Average Delays per Vehicle at each intersection ........................................ 123 Table 34. Case 11 – Results for different combinations of adjustments – values for intersection M at simulation interval 2700 to 3600 ...................................................................................................... 125 Table 35. Sensitivity analysis parameters and tested values. .................................................................... 126 Table 36. List of sensitivity analysis cases and respective tuning algorithm tested. ................................ 127 Table 37. Offset parameter combinations for sensitivity analisys. ........................................................... 127 Table 38. Mainline average delay per vehicle (safety disabled) – eastbound through ............................. 127 Table 39. Mainline average delay per vehicle (safety disabled) – westbound through ............................ 128 Table 40. Split parameter combinations for sensitivity analisys ............................................................... 128 Table 41. Mainline average delay per vehicle (safety disabled) – eastbound left-turn ............................. 130 Table 42. Mainline average delay per vehicle (safety disabled) – westbound left-turn............................ 130 Table 43. Phase Sequence parameter combinations for sensitivity analisys............................................. 131 Table 44. Mainline average delay per vehicle (safety disabled) – eastbound through ............................. 131 Table 45. Cycle length parameter combinations for sensitivity analisys. ................................................. 132 Table 46. Mainline average delay per vehicle (safety disabled) – eastbound through ............................. 133 Table 47. Mainline average delay per vehicle (safety disabled) – westbound through ............................ 134 Table 48. Bell road TOD pattern schedule ............................................................................................... 140 Table 49. Bell Road/Bullard Avenue – 15-Minute MOEs (6000 s to 6900 s) .......................................... 143 Table 50. Bell Road/Litchfield Road - 15-Minute MOEs (6000 s to 6900 s) ........................................... 144 Table 51. Bell Road/Reems Road - 15-Minute MOEs (6000 s to 6900 s)................................................ 144 Table 52. DC Grid SIL- TOD pattern schedule ........................................................................................ 152

xiii

Table 53. 15-Minute MOEs for 14th Street, E Street and Pennsylvania Avenue NW. ............................ 154

xiv

LIST OF ACRONYMS AND ABBREVIATIONS AADT

Average annual daily traffic

ACS

Adaptive control system

ADT

Average daily traffic

ASC MIB

Actuated signal controller management information base

ATMS

Advanced Traffic Management Systems

CCD

Central composite design

CEP

Conflict ending point

CICAS

Cooperative Intersection Collision Avoidance System

CMF

Crash modification factor

CORSIM

Corridor Simulation

CSP

Conflict starting point

DCS

Detection Control System

DeltaS

Maximum speed differential

DeltaV

Change between conflict velocity

DR

Deceleration rate

DOE

Design of experiments

E

East

E/P

Exclusive Permissive

FHWA

Federal Highway Administration

FRESIM

Integrated Traffic Simulator

HGV

High goods vehicle

HOV

High occupancy vehicle

HUTSIM

Helsinki Urban Traffic Simulation

ITE

Institute of Technical Engineers

M

Middle xv

MaxD

Maximum deceleration rate

MaxS

Maximum speed of vehicle

OE

Measures of effectiveness

N

North

OPAC

Optimization Policies for Adaptive Control

PET

Post encroachment time

PHF

Peak-hour factor

PI

Performance Index

PhU

Phase utilization

RHODES

Real Time Hierarchical Optimized Distributed Effective System

SCATS®

Sydney Coordinated Adaptive Traffic System

SCOOT

Split Cycle Offset Optimization Technique

SPF

Safety performance function

SPUI

Single-point urban interchanges

S

South

SCJ

Signal control junction

SIL

Simulation-in-the-loop

SSAM

Surrogate Safety Assessment Methodology

TEXAS

Traffic Experimental Analytical Simulation

TOD

Time-of-day

TRANSIMS Transportation Analysis and Simulation System TTC

Time to collision

V/C

Volume to capacity

W

West

Y+AR

Yellow plus all red

xvi

EXECUTIVE SUMMARY This research comprised two phases: Phase 1 was completed and published in August 2010 under publication FHWA-HRT-10-038. Phase 1 examined the relationships between signal timing and surrogate measures of safety, namely the frequency of rear-end, angle and lanechange conflicts. The FHWA Surrogate Safety Assessment Methodology (SSAM) was used to evaluate various simulated scenarios to test the relationships between signal timing parameters such as cycle, offset, split, phase change interval, detector extension time, left-turn phase protection options and left-turn phase sequence and the occurrence of traffic conflicts. Phase 1 outlined concepts of algorithms to be developed under the Phase 2 research to enhance efficiency and safety. The objectives of this Phase 2 research are to develop real-time adaptive signal timing algorithms to enhance efficiency, and also development of a safety performance function to quantify surrogate safety measures, i.e. number of conflicts, associated with optimized signal timing settings such as cycle, splits, offsets, and also phase sequence. It is the goal of this project to develop a multi-objective optimization methodology using the four principle algorithms that comprise the proposed adaptive system for tuning the cycle length, splits, offsets, and left-turn phase sequence at signalized intersections. The surrogate safety performance function (SPF) is not an optimizer like the efficiency Model, but rather a performance index of surrogate safety measures that is derived based on the selected signal timing settings. When the SPF is enabled, the new signal timing settings will be implemented in the controller only if the safety performance function derives a lesser number of conflicts than with the existing signal timing plans. Otherwise, the current signal timing settings will be retained. If the SPF is disabled, then all new optimized signal timing settings will be implemented in the controller. The surrogate safety measures are represented with a surrogate safety performance function, developed to predict the likely changes in conflicts (measured in SSAM) based on changes to the traffic state and choices for alternative signal timing parameters. Based on the number of traffic state inputs and traffic control variables, the possible number of combinations that could be tested is extremely large. Therefore, all possible combinations cannot be tested, so a statistical methodology called a Design of Experiments (DoE) was used. After specifying a small test network, realistic ranges for input parameters were derived and then generated randomized combinations of test cases using a “Latin Hypercube” approach (Ye, Ko., 1998). Those scenarios are then run in the VISSIM traffic simulation model to obtain efficiency performance (represented by the V/c ratio as reported by the simulation model for each phase at intersection M) and the safety performance (as represented by the total and component totals of traffic conflicts). This information also was used to train the neural network and test it on additional scenarios to see how well the prediction works. The surrogate safety performance function was developed by training a cascade feed forward neural network to learn the relationship between the signal timing settings, efficiency measures, and the resulting average traffic conflict rates. The average traffic conflict rates were postprocessed using SSAM from trajectory data from the VISSIM traffic simulation model. From training the network with approximate 150 scenarios, the average error on the prediction of traffic conflicts on cases that were not used for training was 17%. MATLAB was used to implement the neural network and export the network calculations as a C++ DLL that was integrated with the rest of the adaptive system code. 1

Four principle algorithms and four optimization stages were developed for tuning splits, offsets, cycle time, and phase sequence. The four optimization stages for all algorithms are executed independently, but in sequence and with the feedback steps as shown below. First, (Step 1) the split re-allocation algorithm is executed for each intersection in the system. This identifies if any slack green time can be shifted from one or more phases to another, within the current cycle time, to minimize the maximum phase utilization at the intersection. The safety surrogate is evaluated by checking that the reallocation either provides a benefit by reducing total conflicts or that the reallocation does not exceed a prescribed threshold. After this re-allocation, the offset adjustment algorithm is executed (Step 2) to identify any modifications to the offsets to improve progression. The total percent of arrivals on green, at both the subject intersection and its neighbors, is calculated to represent the efficiency of the proposed change. Similarly to the split calculation, the safety is evaluated by checking that the new offset provides a safety benefit by reducing total conflicts. After the splits and offsets are calculated, modifications to the left-turn phase sequence (Step 3) are evaluated with the new split values calculated in Step 1. If any phase sequence modifications are identified that adjust the offset (the departure platoons to adjacent intersections), the offset calculation is re-executed to determine if this change is of further benefit and can be retained, or if the change is detrimental to performance (efficiency or surrogate safety measures). Finally, the cycle time adjustment algorithm is evaluated (Step 4). Since cycle time affects all of the intersections in the system, it is important that this adjustment is calculated last, after all of the adjustments/improvements to the individual locations are calculated. The real-time adaptive efficiency and surrogate safety algorithms platform developed under this contract is designed to operate in a centralized/distributed signal system. Its execution was emulated and tested in this project with software Simulation-in-the-Loop (SIL). The algorithms require second-by-second polling of all signal controllers, specifically signal timing and detector inputs. Primary measures of performance include: 

Phase utilization. Phase utilization is a surrogate measure of efficiency that represents the degree of saturation of a traffic phase. This measure can be derived directly from the occupancy data measured at stop bar detectors. This measure is used for cycle tuning, cycle selection, split tuning, and phase sequence.



% arrivals on green. % arrivals on green are a measure of efficiency that represents the progression performance of coordinated phases. % arrivals on green can be derived directly from the data measured at upstream detectors on the coordinated phases, and thus requires advanced detectors at 200 feet and preferably 300 feet to work most accurately.



Estimated traffic conflict rate. Total estimated conflicts per hour are a surrogate measure that represents the estimated effect of changing a traffic control parameter on the intersection safety. This measure is a regression model using a feed-forward neuralnetwork that is trained to learn the relationships between signal timing settings and the crossing, lane-changing, and rear-end conflict rates.

Testing and verification analyses performed demonstrated that the tuning algorithms for cycle selection, offsets, splits and phase sequence are promising and deserve further research and 2

application to various flow regimes, in arterial and grid networks, to verify the operation of the various parameter-tuning algorithms under various traffic conditions and roadway configurations. The simulation tests successfully validated the concept of operations of all algorithms, and the derivation of the surrogate safety performance function. The sensitivity analysis also demonstrated that further research is still needed to develop better guidelines on how values within the ranges of the various signal timing parameters should be selected by the user. Kadence, named after the system developed in this research, performed best when all algorithms are enabled in the system parameters set-up: cycle selection, phase sequence, splits and offsets. Enabling one algorithm alone, while disabling the other algorithms, proved to be not effective. Likewise, enabling one or multiple algorithms for only one intersection in a coordinated system was not effective either in producing better signal timings than pre-optimized timing settings. The surrogate safety performance function was validated successfully. Kadence predicts the total number of conflicts associated with optimized signal timing settings. However, the more efficient timing settings did not always result in a lesser number of conflicts than the preoptimized signal timing settings. Therefore, users should be aware that the Safety Performance Function is not an optimizer of safety surrogate measures but rather an absolute index representing the total number of conflicts derived from the signal timings. This project proved to be very complex and demanded in-depth understanding of the algorithms and the set-up of the parameters in Kadence. Users must be aware that adaptive signal timing strategies are not always better than time-of-day generated signal timing plans unless the user is intimately familiar with the set-up and operation of the adaptive algorithms, and also is able to implement adaptive signal timing under appropriate flow regimes, i.e. unsaturated traffic flow conditions. All traffic flow regimes tested in this project were unsaturated conditions with volume to capacity ratio in the range of 0.60 to 1.0. A few saturated traffic flow conditions, with volume-to-capacity ration of 1.2 and greater, were tested and the results were not better than preoptimized signal timing settings.

3

4

1 INTRODUCTION This research, under a SBIR phase 2, focuses on the development of real-time adaptive signal timing algorithms to enhance efficiency, and also development of a safety performance function to quantify surrogate safety measures, i.e. number of conflicts, associated with optimized signal timing settings such as cycle, splits, offsets, and also phase sequence. The real-time algorithms consist of cycle selection, split, offset and phase sequence methodologies that operate in real-time under a centralized/distributed signal system control. This research comprised two phases: Phase 1 was completed and published in August 2010 under publication FHWA-HRT-10-038 (Sabra, et. al. 2010). Phase 1 examined the relationships between signal timing and surrogate safety measures, namely the frequency of rear-end, angle and lane-change conflicts. The FHWA Surrogate Safety Assessment Methodology (SSAM) was used to evaluate various simulated scenarios to test the relationships between signal timing parameters such as cycle, offset, split, phase change interval, detector extension time, left-turn phase protection options and left-turn phase sequence and the occurrence of traffic conflicts. 1.1 SUMMARY OF PHASE I RESEARCH RESULTS The Phase 1research examined the relationships between signal timing and surrogate measures of safety. The Federal Highway Administration (FHWA) Surrogate Safety Assessment Methodology (SSAM) was used to evaluate simulated scenarios to test the relationships between signal-timing parameters and the occurrence of traffic conflicts. Only a single intersection and a three-intersection arterial were examined, and each parameter was tested independently. The analysis effort indicates the following results: 

The ratio of demand to capacity (i.e., the length of the split) is a factor that influences the total number of conflicts. There is an inverse linear relationship between splits and total conflicts.



Cycle length has the most significant impact on the total number of conflicts.



Detector extension times have only a minor impact on changes to conflict rates.



The phase-change interval has a marginal effect on the total number of conflicts.



Left-turn phasing (protected/permitted) has a significant effect on the total number of conflicts.



An offset has insignificant effect on conflicts until the change is more than ±10 percent of the cycle length.



Phase sequence has a significant effect on the total number of conflicts on an arterial.

These results were obtained by modifying each variable independently for specific geometric and volume conditions. As such, these results provided evidence that certain parameters have a positive correlation to changes in surrogate measures of safety, but they do not provide metrics that can be used for real-time signal-timing optimization. The Phase 1 research also discussed a methodology based on design of experiments to calculate a safety performance function that can be used for estimating the effect of changes to signal-timing parameters in tandem. The report concluded with the development of a multi-objective optimization methodology and five 5

principle algorithms that constitute the proposed adaptive system for tuning cycle length, splits, offsets, left-turn phase protection treatment, and left-turn phase sequence of a set of intersections. 1.2 RESEARCH OBJECTIVES The objective of this research is to develop a multi-objective optimization methodology using the four principle algorithms that comprise the proposed adaptive system for tuning the cycle length, splits, offsets, and left-turn phase sequence at signalized intersections. A second objective is to develop of a safety performance function to quantify surrogate safety measures associated with optimized signal timing settings such as cycle, splits, offsets, and also phase sequence. The surrogate safety performance function (SPF) is not an optimizer like the efficiency Model, but rather a performance index of surrogate safety measures that is derived based on the selected signal timing settings. When the SPF is enabled, the new signal timing settings will be implemented in the controller only if the safety performance function derives a lesser number of conflicts than with the existing signal timing plans. Otherwise, the current signal timing settings will be retained. If the SPF is disabled, then all new optimized signal timing settings will be implemented in the controller. The work plan for research was developed under the following activities: 1.2.1 Develop the Design of Experiments (DoE) In this task, the designed experiment was specified, and 150 simulation cases of the designed experiment were run to collect the necessary traffic conflict data resulting from the impacts of interactions between the input parameters. Approximately 10% of the runs were then set aside for the theoretical validation process as discussed below. The remaining runs are then used to calculate the regression coefficients of the safety performance function. Several functional forms of the regression equation were tested according to the results of the regression process. 1.2.1.1 Develop and test the safety performance function The predictive performance of the safety function is tested with the ~10% of the simulation runs that were held out of the regression process. This step was important to identify the ability of the safety performance model to closely replicate the data obtained directly from running a particular simulation case. 1.2.2 Develop the multi-objective adaptive algorithms In Task 3, the adaptive control algorithms were implemented in software. These algorithms require both detector and phase timing data. These data were obtained from a microscopic simulation model. Data processing and evaluation concepts used in this research are based on FHWA’s earlier effort of traffic adaptive algorithm development for split and offset tuning and add the evaluation of the safety performance function and the additional tuning steps as outlined in later sections. The other algorithms for phase sequence and cycle length selection are a new development under this project including a new communications processing component that is developed to avoid any proprietary issues with earlier efforts performed by FHWA.

6

1.2.2.1 Evaluate the performance in off-line scenarios and implement the algorithms in a realtime processing system Micro-simulation data were obtained for several test cases and output to files. The data files were input to the adaptive control calculation engine to verify that the compromise algorithm process executes as expected. Since there is no feedback loop to the simulation process, the results of the changes to the traffic signal parameters cannot be assessed for the “real” performance, so this task simply verifies and demonstrates that the algorithms are functioning. The algorithms are implemented in an on-line manner with a traffic simulation model in the loop. To streamline the process of moving the algorithms to real-world testing, the algorithms were interfaced to a virtual traffic controller that is identical to the version of the controller software that runs in the field. There are two such controllers available in the market today – D4 from 4th Dimension Traffic and the Econolite ASC/3. Both are able to interface to the VISSIM microsimulation model environment. The research team used the D4 virtual controller software. In either case, the interface between the optimization component and the field controller will be an open standard, such as NTCIP. This implementation allows evaluation of the performance of the algorithms before deployment in the street. A similar architecture is envisioned for the system when deployed in the real world. The adaptive control algorithms will reside as part of a central system similar to the architecture of the lite version of ACS by polling data from the controllers (detector data and phase timing information) and downloading new timing parameters on a periodic basis (i.e. each 3-5 cycles). This adaptive control component will be part of a central system, unlike the lite version of ACS, which operates as a master controller. 1.2.2.2 Test the algorithms’ performance for real-life simulation scenarios Algorithms were tested with simulation scenarios. A combination of scenarios for single intersections, small arterial, and a small grid network were tested. The test cases were selected to replicate field locations that could become candidates for future field testing. 1.3 ORGANIZATION OF THE REPORT This report is organized into the following six sections: 

Section 1: provides an overview of the project and a statement of the research objectives.



Section 2: discusses the experimental design for the development of the safety performance function.



Section 3: discusses the development of the safety performance function.



Section 4: Discusses the development of multi-objectives algorithms for cycle length, spilt, offset and phase sequence tuning.



Section 5: defines the system parameters and system user interface used to program each of the tuning algorithms.



Section 6: discusses the results of various test case scenarios used in a software and hardware simulation-in-the-loop environment to test the operation of the various tuning 7

algorithms, interface with the virtual controller, interface with the adaptive system, and also to verify the execution of the algorithms. 

Section 7: discusses the results of real-life arterial and grid network applications used to verify and validate the results of the algorithms.



Section 8: identifies and suggests future research enhancements to the adaptive algorithms.

8

2 EXPERIMENTAL DESIGN 2.1 EXPERIMENTAL DESIGN FOR THE DEVELOPMENT OF A SAFETY PERFORMANCE FUNCTION In Phase 1 of this research, the relationships between traffic signal timing parameters and safety performance were explored. Several test cases were constructed to identify isolated correlation effects between changes to one signal timing parameter (e.g. cycle time) and the relative change in safety performance. Surrogate safety performance in Phase 1 was estimated using the FHWA Surrogate Safety Analysis Methodology (SSAM) tool and the VISSIM traffic micro-simulation model (Gettman, et. al. 2012 and PTV AG, 2008). Analysis results in Phase 1 indicated that there are measurable and statistically significant differences in the generation of conflict events when modifying several of the signal timing parameters in the test cases (Sabra, et. al. 2010). The preliminary analysis performed with SSAM to date has established that certain parameters have positive correlations with changes in safety, but this analysis cannot be used directly to determine the magnitude of effects. In addition, each parameter was evaluated in isolation. In a real-world, real-time setting, changes to these parameters must be evaluated simultaneously and in an integrated fashion. Prediction of changes to traffic efficiency is somewhat straightforward and has been proven effective in existing adaptive control systems including the ACS lite, RHODES, OPAC, SCOOT, and SCATS, among others. Prediction of changes to traffic safety has two primary challenges: the choice of the measure or measures to be predicted, and the functional form. Since conflicts were determined in Phase 1to have statistically significant differences when traffic signal timing parameters are modified (as estimated by simulation models), we used the prediction of traffic conflicts. For the functional form, we chose a neural network. Neural networks are very good at embedding pattern recognition relationships in nonlinear ways. This approach is similar to traditional crash prediction, except that the prediction domain is sufficiently higher fidelity than predicting the crashes that might occur over the next year. The safety performance function evaluates how the conflict occurrence rates would be perturbed over the next hour if the input traffic demands and signal timing control parameters are set. In this approach, we establish a function: Safety performance index = f(signal parameters, traffic characteristics) The safety performance “index” is defined as the total number of conflicts at the intersection. The signal parameters that are input to the function are a combination of the traffic signal control parameters (cycle, splits, offsets, phase sequence), road geometry (primarily distance between intersections), and the real-time state of the traffic (degree of saturation, queue lengths). Based on the number of traffic state inputs and traffic control variables, the possible number of resulting combinations is extremely large. All possible combinations cannot be tested to build a relationship between the inputs and the outputs, so we used a statistical methodology called a Design of Experiments (DoE) to sample states and control settings from the set of all possible combinations in order to get a representative sampling. There are several methods for DoE. In Phase 1, we postulated the use of a Central Composite Design (CCD). This method assumes linear and quadratic relationships between input variables and output data (traffic conflicts). 9

After further discussion, we decided to use a “Latin Hypercube” approach. This approach is a randomization strategy that attempts to cover the solution set of the “Hypercube” of possible input variables. This approach allows for non-parametric and more flexible choices for functional forms which we decided was necessary since the functional form of the safety prediction function is a neural network. Since the adaptive control approach selected for our adaptive algorithm is a parameter tuning approach, only small changes to signal timing input variables are made at each optimization step. Since the efficiency improvement search algorithms are quite simple, the safety performance function also only needs to be able to make predictions of marginal differences between settings. In addition, since the adaptive parameter tuning approach considers the optimization of each intersection independently, we decided that the safety performance calculator can predict the performance of the system for each intersection independently. Thus, a relatively simple network could be used for generation the test cases for training the neural network. To this regard, we have devised a two-stage experimental design. In the first stage, a reasonably small number of characteristic volumes, route proportion, and geometric conditions will be generated. Then, for each of these scenarios, an optimal set of signal timings is calculated using the microscopic analysis tool Synchro (Trafficware, 2012). From those baseline cases, perturbations are made so that the traffic situation is no longer correctly matched to the signal timing parameters. The perturbations are chosen using the Latin Hypercube theory to result in approximately 150 test scenarios. Those scenarios are run through the simulation model to generate traffic performance and safety performance results. A simple network (similar to that used in Phase 1) was used as the basis for the test cases. This test network is shown in Figure 1. Link N-M

Link M-W

Int W

Link E-M

Int M

Int E

Link M-E

Link W-M

Link for SSAM Analysis

Link S-M

Figure 1. Illustration. Test network for experimental design Three intersections are necessary for testing the response of changes to offsets and sequence from and to neighboring intersections. For simplicity, the intersections are named “M”, “E”, and “W”, for “Middle”, “East”, and “West”, respectively. As shown in Figure 2, only the green shaded links were used for calculating both efficiency and safety performance results. This 10

focuses the performance on all of the approaches to the subject intersection M and the departure links from M to E and W. In order to reduce the combinations of possible test scenarios, we made some additional assumptions about which inputs and parameters to modify and which to leave fixed. Figure 2 illustrates which volume inputs were modified from the baseline settings and which were left the same. Blue arrows indicate which volumes were modified and red arrows indicate which volumes remained the same. Link N-M

Link M-W

Int W

Link E-M

Int M

Int E

Link M-E

Link W-M

Link for SSAM Analysis

Link S-M Variable arrival volume Fixed arrival volume

Figure 2. Illustration. Variable and fixed arrival volumes Figure 3 illustrates which route proportions were changed and which were set to a constant proportion. Vehicles in VISSIM follow routes that they are assigned to at input points and remain on that route until they reach the end of the route. At that point they choose the next route according to a route distribution or they exit the network. In order to make the traffic flows realistic, in our small network we included routes at each input point that continue through the entire system. As shown in Figure 3 , for example, there are five possible routes from a side street entrance at W that principally make a right turn to contribute to the performance evaluation: 

Straight through and exit



Left turn and exit



Right turn and continue straight until exit



Right turn and then left turn at M and exit



Right turn and then left turn at E and exit 11

Link for SSAM Analysis

Fixed proportion Variable proportion

Figure 3. Illustration. Right turn routes The identical (but where E is replaced with W) route distribution is applied at the southbound right turn at E. As noted in the figure, red arrows indicate where the route proportion was kept fixed and blue arrows indicate where the route proportion was changed to create different travel patterns. For the East and West bound inputs at E and W the route distributions that were considered are shown in Figure 4. As shown in Figure 4, for example, there are six possible routes from the through entrance at W: 

Straight through and exit



Left turn and exit



Right turn at M and exit



Right turn at E and exit



Left turn at M and exit



Left turn at E and exit

The identical (but where E is replaced with W) route distribution is applied at the through movement at E.

12

Link for SSAM Analysis

Fixed proportion Variable proportion

Figure 4. Illustration. Through routes For the Left turn inputs at E and W the route distributions that were considered are shown in Figure 5. As shown in Figure 5, for example, there are five possible routes from the through entrance at W: 

Straight through and exit



Right turn and exit



Left turn and the right turn at M and exit



Left turn and the right turn at E and exit



Left turn and exit straight through

13

Link for SSAM Analysis

Fixed proportion Variable proportion

Figure 5. Illustration. Left turn routes The identical (but where E is replaced with W) route distribution is applied at the left turn movement at E. Finally the side street approaches at M as shown in Figure 6 were distributed as follows for the Northbound approach: 

Straight through and exit



Right turn and exit



Left turn and exit



Right turn and left turn at E to exit



Left turn and right turn at M to exit

The identical (but where E is replaced with W) route distribution is applied at the southbound approach.

14

Link for SSAM Analysis

Fixed proportion Variable proportion

Figure 6. Illustration. Side street approach routes These changes, along with the modifications to volumes, generates a significant set of traffic scenarios inputs were modified from the baseline settings and which were left the same. From a geometry perspective, the only variable that was adjusted was the distance between the intersections, as illustrated in Figure 7.

Link for SSAM Analysis

Variable distance

Figure 7. Illustration. Link lengths modified in experimental design From a control perspective, we decided to fix the operation of intersections E and W and vary the splits from the optimized settings only at M. Our goal was not to create oversaturated or 15

significantly under-saturated conditions at E and W, as this may skew the resulting conflict and traffic efficiency measures at M on the links of interest. This is illustrated in Figure 8.

25% 25%

12.5% 12.5% 12.5% 12.5%

25% 25%

25% 25%

12.5% 12.5% 12.5% 12.5%

25% 25%

Link for SSAM Analysis

25% 25% 25% 25% Webster fixed time splits

12.5% 12.5% 12.5% 12.5% 12.5% 12.5% 12.5% 12.5%

Variable Splits

Figure 8. Illustration. Splits adjusted in the experimental design Cycle times were adjusted together at all three intersections and kept at the same value. Splits were adjusted proportionally according to the percentage allocation in the original solution. Sequence and left turn type was only modified at intersection M in any scenario. This seemed reasonable since the effects of modified sequence or left turn type at E and W would probably be overshadowed by the effects of offset adjustments and changes in flow rates and route proportions at E and W. Offsets were adjusted at W and E but not at M. This is illustrated in Figure 9. This also seemed reasonable since a combination of changes at M and E could be accomplished by a change only at E. For example if offset at E was changed from 10s to 15s and offset at M was perturbed from 0s to 10s (a relative change from 10s to 5s) the same can be achieved by changing the offset at E from 10s to 5s. Thus the offset at M was always set to 0s.

16

Link for SSAM Analysis

Variable Offset Fixed Offset

Figure 9. Illustration. Offsets adjusted in the experimental design. These input factors were then modified to span the ranges and values identified in Table 1.

17

Table 1. Input factors and ranges for a three-intersection model: W (west), M (middle), E (east) INPUT FACTOR/ VARIABLE Intersection W NB Traffic Volume

SB traffic volume EB traffic volume NB routes Right to EBLT at M Right to EBT at M and E Right to EBLT at E NBT NBL SB routes Left to EBRT at M Left to EBT at M and E Left to EBRT at E SBT SBR EB routes Through to EBL at M Through to EBR at M Through to EBL at E Through to EBT at M and E Through to EBRT at E EBL Intersection M NB Traffic Volume SB traffic volume NB routes Left to WBRT at W Left to WBT at W Right to EBLT at E NBT Right to EBT at E SB routes Left to EBRT at E Left to EBT at E Right to WBLT at E SBT

RANGE

800 to 1800

REMARKS

** sanity checks to make sure all volumes of particular scenario do not create serious gridlock at intersection M

800 to 1800 1200 to 2400 5%, 10%, 15%, 20% 30%, 25%, 20%, 15% 5% 55% 15%

Coupled with next variable

5%, 10%, 15%, 20% 20%, 15%, 10%, 5% 5% 55% 5%

Coupled with next variable

5%, 10%, 12%, 14% 10%, 10%, 7%, 2% 25%, 15%, 6%, 4% 45%, 55%, 65%, 70%

Coupled with next 3 variables

5% 5%

Fixed Fixed

Fixed Fixed Fixed

Fixed Fixed Fixed

800 to 1800 800 to 1800 5% 35%, 30%, 25%, 20% 5% 25%, 35%, 45%, 55% 30%, 25%, 20%, 15%

Fixed Coupled with other three variables Fixed

5% 35%, 30%, 25%, 20% 5% 25%, 35%, 45%, 55%

Fixed Coupled with other three variables Fixed

18

INPUT FACTOR/ VARIABLE Right to WBT at E Intersection E NB Traffic Volume SB traffic volume WB traffic volume SB routes Right to WBLT at M Right to WBT at M and W Right to WBLT at W SBT SBL NB routes Left to WBRT at M Left to WBT at M and W Left to WBRT at W NBT NBR WB routes Through to WBL at M Through to WBR at M Through to WBL at W Through to WBT at M and W Through to WBRT at W WBL Intersection W NBT Number of Lanes NB RT Number of Lanes NB LT Number of Lanes SBT Number of Lanes SB RT Number of Lanes SB LT Number of Lanes EBT Number of Lanes EB RT Number of Lanes EB LT Number of Lanes WBT Number of Lanes WB RT Number of Lanes WB LT Number of Lanes Intersection M NBT Number of Lanes NB RT Number of Lanes NB LT Number of Lanes SBT Number of Lanes

RANGE

REMARKS

30%, 25%, 20%, 15% 800 to 1800 800 to 1800 1200 to 2400 5%, 10%, 15%, 20% 20%, 15%, 10%, 5% 5% 55% 15%

Coupled with next variable

5%, 10%, 15%, 20% 20%, 15%, 10%, 5% 5% 55% 5%

Coupled with next variable

5%, 10%, 12%, 14% 10%, 10%, 7%, 2% 25%, 15%, 6%, 4% 45%, 55%, 65%, 70%

Coupled with next 3 variables

5% 5%

Fixed Fixed

3 1 1 3 1 1 3 1 1 3 1 1

Fixed Fixed Fixed

Fixed Fixed Fixed

Storage will facilitate Max. Queue

3 0, 1 1, 2 3 19

INPUT FACTOR/ VARIABLE SB RT Number of Lanes SB LT Number of Lanes EBT Number of Lanes EB RT Number of Lanes EB LT Number of Lanes WBT Number of Lanes WB RT Number of Lanes WB LT Number of Lanes Intersection E NBT Number of Lanes NB RT Number of Lanes NB LT Number of Lanes SBT Number of Lanes SB RT Number of Lanes SB LT Number of Lanes EBT Number of Lanes EB RT Number of Lanes EB LT Number of Lanes WBT Number of Lanes WB RT Number of Lanes WB LT Number of Lanes Spacing between Intersections Spacing E to M Spacing M to W Cycle Length Splits – intersection E Splits – intersection W Splits – intersection M Split- Phase 1 (WBL)

Split- Phase 2 (EBT) Split- Phase 3 (SBL) Split- Phase 4 (NBT) Split- Phase 5 (WBL) Split- Phase 6 (WBT) Split- Phase 7 (NBL) Split- Phase 8 (SBT) Sequence - Intersection M NB Left Turn Sequence SB Left Turn Sequence

RANGE

REMARKS

0, 1 1, 2 3 0, 1 1, 2 3 0, 1 1, 2 3 1 1 3 1 1 3 1 1 3 1 1

Storage will facilitate Max. Queue

400, 600, 800, 1000, 1200, 1600, 2400 400, 600, 800, 1000, 1200, 1600, 2400 60, 90, 120, 180, 240 Fixed time Solve via Synchro Fixed time Solve via Synchro +/- 5%, 10%, 15%

+/- 5%, 10%, 15%

+/- 5%, 10%, 15% +/- 5%, 10%, 15%

Need sanity checks on these combinations for oversaturation combined with volumes/routes Coupled with splits 1,3,4 Coupled with splits 1,2,4 Coupled with splits 1,2,3 Coupled with splits 6,7,8 Coupled with splits 5,7,8 Coupled with splits 5,6,8 Coupled with splits 5,6,7 “none” only if LTT = “permissive” “none” only if LTT = “permissive”

None, Lead, Lag None, Lead, Lag 20

INPUT FACTOR/ VARIABLE EB Left Turn Sequence WB Left Turn Sequence Left turn – intersection M NB Left Turn Treatment SB Left Turn Treatment EB Left Turn Treatment WB Left Turn Treatment Offsets Offset W

Offset E

Offset M Global Values Minimum Green, Left Minimum Green, Through Y+AR Detector Gap Extension Pedestrian Timing Main line speed Side Street Speed Left Turn Speed Right Turn Speed Lane Width Grade Storage Lanes

RANGE

REMARKS “none” only if LTT = “permissive” “none” only if LTT = “permissive”

None, Lead, Lag None, Lead, Lag Permissive, Protected, Permissive/protected Permissive, Protected, Permissive/protected Permissive, Protected, Permissive/protected Permissive, Protected, Permissive/protected 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90% 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90% 0

None None None 30, 35, 40, 45, 50 30, 35, 40, 45 15 mph 9 mph 12 0% 250ft if link > 600ft; 150 ft. if link < 600 ft.

Percent

Fixed Reference setting Default to 5 seconds Default to 10 seconds Fixed Per ITE Criteria Fixed per Dilemma Zone Guidelines Not included/embedded in Splits

Fixed Fixed Fixed Fixed

150 combinations from the variables listed in the above table were generated using the Latin Hypercube randomization approach. A program was developed that generated the VISSIM files automatically when the geometry and factors in the ring-barrier controller (RBC) were modified according to the parameters above. The summary of the control parameters of the first 100 test cases is listed in Table 2. Care was taken to identify and remove scenarios where the system became overwhelmingly oversaturated. In such situations, neither the efficiency nor the safety performance metrics are reliable. Efficiency metrics such as v/c ratios are capped at 1.0 and total conflicts increase nonlinearly in an unrealistic manner in the simulation model.

21

Spacing M to W

2400 2400 600 400 1600 1200 600 600 1200 1000 1000 400 800 1600 1600 1000 1000 1000 1000 400 400 600 800 400 2400 2400 2400 2400 800 1200 400 600 800 800 1200 1200 1600 1600 600

SCN

1_A 1_B 3_B 5_A 7_C 12_B 16_B1 16_B2 19_C 22_B 22_C 25_B 31_B 38_B1 38_B2 43_B1 43_B2 46_B 46_C 54_A 54_C 55_B 56_B 62_B 69_A 69_B 69_C 72_B 77_B 87_C 89_C 94_A 96_C 99_C 100_B 100_C 104_B 104_C 105_A

Spacing E to M

1200 1200 2400 800 1200 600 1600 1600 1600 600 600 1200 800 1600 1600 1000 1000 2400 2400 1600 1600 400 1200 1000 2400 2400 2400 800 1000 800 600 2400 2400 1600 1000 1000 2400 2400 800

Split- Phase 1 (WBL)

15.1 14.0 14.0 14.5 13.0 14.0 12.3 16.8 20.2 12.7 12.4 19.1 11.5 16.0 10.5 14.5 17.0 13.7 20.2 10.6 11 11.5 10.5 9 13.9 10.5 17.9 21.4 14.7 17.1 16 11.5 11.1 28.7 11.4 22 20.2 11.5 13.5

Split- Phase 2 (EBT) 24.2 30.0 26.0 23.0 55.6 28.5 32.2 18.7 34.8 27.8 26.6 25.8 19.9 45.5 27.5 38.0 21.0 48.2 52.7 24.9 45.5 19.7 31.5 20 17.1 16 37.1 46.6 18.7 52.5 28 25.1 24.4 42.4 28.5 36 18.9 27.5 17.5

Split- Phase 3 (NBL) 13.5 11.0 17.0 16.0 16.6 10.5 12.1 20.8 17.0 13.7 13.3 19.2 15.4 10.5 15.0 15.0 15.0 12.9 11.8 10.5 10.5 13.2 11 12 13.5 10.5 15 16 14.4 15.4 14 11.5 15.5 15.2 14 13 18 17.7 13.5

Split- Phase 4 (SBT) 17.2 25.0 13.0 16.5 24.8 17.0 23.4 13.7 18.0 15.8 37.7 15.9 23.2 18.0 17.0 12.5 17.0 15.2 25.3 14 23 15.6 17 19 15.5 23 30 16 22.2 25 22 21.9 19 23.7 16.1 19 12.9 23.3 15.5

Split- Phase 5 (EBL) 17.4 16.0 10.5 14.5 19.1 11.0 12.0 11.0 15.0 17.2 14.0 19.2 12.7 21.4 12.1 19.0 17.0 21.0 27.5 11 14 11.5 15 10.5 13.8 10.5 15 18 10.5 19.7 17.5 18.9 11.2 15.7 14 14.1 11 22 14.2

Split- Phase 6 (WBT) 21.9 28.0 29.5 23.0 49.5 31.5 32.5 24.5 40.0 23.3 25.0 25.7 18.7 40.1 25.9 33.5 21.0 40.9 45.4 24.5 42.5 19.7 27 18.5 17.2 16 40 50 22.9 49.9 26.5 17.7 24.3 55.4 25.9 43.9 28.1 17 16.8

Split- Phase 7 (SBL) 13.9 14.9 10.5 14.5 14.4 10.5 10.5 10.5 14.1 12.0 14.0 13.3 22.8 10.5 12.2 12.5 15.0 11.1 18.1 10.5 10.5 13.1 15.1 12.4 13.5 10.5 10.9 17.7 10.5 18.4 17 11.5 10.5 18 16 18 16.5 18 13.5

Split- Phase 8 (NBT) 16.8 21.1 19.5 18.0 27.0 17.0 25.0 24.0 20.9 17.5 37.0 21.8 15.8 18.0 19.8 15.0 17.0 17.0 19 14 23 15.7 12.9 18.6 15.5 23 34.1 14.3 26.1 22 19 21.9 24 20.9 14.1 14 14.4 23 15.5

Cycle Length

22

70 80 70 70 110 70 80 70 90 70 90 80 70 90 70 80 70 90 110 60 90 60 70 60 60 60 100 100 70 110 80 70 70 110 70 90 70 80 60

Offset W 33 48 5 22 105 33 5 31 38 41 9 49 5 63 1 12 12 3 96 37 0 59 34 10 31 26 92 50 52 6 11 10 35 20 11 12 65 71 26

Offset M 69 3 29 31 108 2 2 36 88 8 1 1 68 18 22 16 19 0 92 30 0 56 0 8 5 0 30 83 0 106 77 38 46 0 10 88 16 4 4

Offset E 64 2 0 26 23 6 47 0 88 0 8 71 0 21 67 59 52 48 48 8 86 33 60 54 0 57 86 80 2 12 8 62 0 1 48 6 67 60 0

3 3 2 3 3 2 2 2 2 2 2 2 2 2 2 3 3 2 2 2 2 2 3 2 2 2 2 3 2 2 3 3 2 2 2 2 3 3 3

EB Left Turn Treatment

Table 2. Control parameters of 100 test cases WB Left Turn Treatment 3 3 2 3 3 2 2 2 2 2 2 2 2 2 2 3 3 2 2 2 2 2 3 2 2 2 2 3 2 2 3 3 2 2 2 2 3 3 3

NB Left Turn Treatment 3 3 2 3 2 3 3 3 2 2 2 2 3 3 3 2 2 2 2 3 3 2 2 2 3 3 3 2 2 2 3 3 2 2 2 2 2 2 2

SB Left Turn Treatment 3 3 2 3 2 3 3 3 2 2 2 2 3 3 3 2 2 2 2 3 3 2 2 2 3 3 3 2 2 2 3 3 2 2 2 2 2 2 2

NB Left Turn Sequence 1 1 1 1 2 1 1 2 2 2 2 1 1 2 1 2 2 2 1 1 1 2 1 2 1 1 1 2 2 2 1 2 2 2 2 2 2 2 2

SB Left Turn Sequence 1 1 1 1 2 1 1 2 2 2 2 1 1 2 1 1 2 2 2 1 1 2 2 1 1 1 1 2 2 2 1 2 2 2 2 2 1 1 2

EB Left Turn Sequence 1 1 1 2 2 1 2 2 1 1 2 1 2 1 2 2 2 2 2 2 1 2 1 2 2 2 2 2 1 2 2 1 2 2 2 2 1 2 1

2 2 2 2 2 2 1 1 2 2 2 2 2 2 1 1 1 1 2 2 2 1 1 2 1 1 2 2 2 1 2 1 1 2 1 2 1 1 2

WB Left Turn Sequence

Spacing M to W

600 1200 400 400 600 2400 2400 400 800 800 600 1200 1200 2400 1600 600 800 400 1200 1600 1200 800 2400 1000 1000 800 1600 1200 600 1600 400 2400 2400 1000 1000 600 600 400 1200 1200 1600 1600

SCN

105_C 107_B 118_B 118_C 122_B 124_A 124_B 136_B 141_B 141_C 147_A 4_B 4_C 6_C 8_C 9_C 10_B 11_B 14_B 15_A 17_B 24_B 26_B 27_B 28_A 29_B 30_C 32_B 33_A 35_B 36_B 37_B 39_B 40_C 41_C 42_A 44_B 47_B 51_B 52_B 53_B 57_A

Spacing E to M

800 2400 800 800 1200 400 400 1000 800 800 600 800 800 600 600 400 400 1000 1000 1600 1000 600 800 1000 800 1000 1000 1000 400 600 1200 800 600 2400 2400 2400 800 800 1200 1200 1600 1600

Split- Phase 1 (WBL)

10.9 14 15.6 12 12.7 11.5 14 23 18 10.9 18 11 14.1 16.2 20.1 10.5 17 18.1 16 12.5 16.2 15.1 12.6 12 15.4 11.9 13 13 16.5 12 12.4 12.5 17 20.5 17 12.5 14 11.5 23 14 12.7 10.5

Split- Phase 2 (EBT) 49.1 25 22.7 17 29.6 22.7 19 20 22 29.4 22 20.5 37.4 43.8 97.9 23.5 26.5 40.9 27.8 26.5 25.7 30.4 18.3 22.9 19.1 18.3 70 21 20.5 27 18.1 33 42 58.5 46.5 19.4 26 33 41 32 20.3 24.2

Split- Phase 3 (NBL) 11.5 14 15.2 11 13.5 11.9 13 12.5 17 12.2 17 15 22 20 20.9 12.1 14 14.4 10.9 18.5 13.7 15.8 16.6 11.5 10.7 11.5 10.5 13 15.5 12 10.5 12 12.7 11 21 12.5 14 11 11 11 12.5 10.8

Split- Phase 4 (SBT) 18.5 17 16.5 30 14.2 13.9 14 14.5 13 27.5 13 13.5 16.5 20 21.1 23.9 12.5 16.6 25.3 12.5 14.4 18.7 12.5 13.6 14.8 18.3 26.5 13 17.5 19 19 12.5 18.3 20 25.5 15.6 16 14.5 25 13 14.5 14.5

Split- Phase 5 (EBL) 12 14 14.1 11 14.1 12.4 13 11.3 13 13.7 13 14 18.5 13.6 21 11 21 16 17 15.2 16.6 14.8 12 16.1 14.3 13.6 17 16.3 15.8 15 13 14 15 19 16.3 12.6 12.5 14.4 18 16 14.3 15.7

Split- Phase 6 (WBT) 48 25 24.2 18 28.2 21.8 20 31.7 27 26.6 27 17.5 33 46.4 97 23 22.5 43 26.8 23.8 25.3 30.7 18.9 18.8 20.2 16.6 66 17.7 21.2 24 17.5 31.5 44 60 47.2 19.3 27.5 30.1 46 30 18.7 19

Split- Phase 7 (SBL) 12 15 10.5 10.7 12.7 11.6 11.2 11 10.5 10.5 10.5 16 24 23 23 12 11 12 22 12.9 12 16 14 11.6 12 11.5 14 13 15.5 15 15.8 12 14 17 15.8 13.6 12.3 10 14.9 11 12.5 11.1

Split- Phase 8 (NBT) 18 16 21.2 30.3 15 14.2 15.8 16 19.5 29.2 19.5 12.5 14.5 17 19 24 15.5 19 14.2 18.1 16.1 18.5 15.1 13.5 13.5 18.3 23 13 17.5 16 13.7 12.5 17 14 30.7 14.5 17.7 15.5 21.1 13 14.5 14.2

Cycle Length

23

90 70 70 70 70 60 60 70 70 80 70 60 90 100 160 70 70 90 80 70 70 80 60 60 60 60 120 60 70 70 60 70 90 110 110 60 70 70 100 70 60 60

Offset W 11 2 32 29 16 36 28 8 8 8 0 34 46 55 28 15 44 8 51 43 31 22 8 36 25 7 21 25 35 39 9 2 54 39 6 53 9 10 40 51 2 34

Offset M 84 0 66 0 7 8 1 65 68 0 56 5 2 0 1 4 8 89 12 12 5 8 4 7 2 23 4 58 2 2 40 2 4 15 2 57 69 0 0 14 38 0

Offset E 79 33 1 2 44 2 6 4 0 3 2 0 80 3 4 8 8 8 8 18 5 0 6 58 0 24 13 56 68 68 1 1 87 88 53 6 7 10 90 18 8 1

EB Left Turn Treatment 3 2 2 2 2 2 2 2 2 2 3 2 2 2 2 2 2 3 2 2 2 2 2 2 3 2 2 2 3 2 3 3 3 2 2 2 2 2 3 3 2 2

WB Left Turn Treatment 3 2 2 2 2 2 2 2 2 2 3 2 2 2 2 2 2 3 2 2 2 2 2 2 3 2 2 2 3 2 3 3 3 2 2 2 2 2 3 3 2 2

NB Left Turn Treatment 2 2 3 3 2 2 2 3 3 3 2 3 2 2 2 2 2 3 3 2 2 2 2 2 2 2 2 3 2 2 3 2 3 2 3 2 3 2 2 2 2 3

SB Left Turn Treatment 2 2 3 3 2 2 2 3 3 3 2 3 2 2 2 2 2 3 3 2 2 2 2 2 2 2 2 3 2 2 3 2 3 2 3 2 3 2 2 2 2 3

NB Left Turn Sequence 1 2 1 1 2 2 2 1 1 1 2 1 1 2 2 2 1 1 1 1 2 2 1 2 2 2 2 1 1 1 2 2 1 2 1 2 1 1 1 2 2 2

SB Left Turn Sequence 2 2 1 1 2 2 2 1 1 1 1 1 1 2 2 2 2 1 1 2 1 2 2 2 2 1 2 1 1 2 2 2 1 1 1 2 1 1 2 2 2 2

EB Left Turn Sequence 2 2 1 1 2 1 2 2 2 2 2 1 1 2 2 2 1 2 1 1 1 1 1 1 1 1 2 1 2 1 1 1 1 2 2 2 2 2 1 1 2 1

2 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 1 2 2 2 2 1 2

WB Left Turn Sequence

Spacing M to W

600 1200 600 1600 2400 2400 600 2400 800 1000 1200 1000 1600 2400 1600 1000 2400 1200 1000 800 800

SCN

61_B 63_A 64_A 66_B 67_C 73_C 74_B 76_B 78_B 81_B 86_A 91_B 108_B 112_B 115_C 130_C 131_B 133_B 137_C 139_A 145_C

Spacing E to M

800 400 2400 2400 600 1600 2400 600 1200 600 800 1200 1200 1200 2400 1600 1600 600 1000 2400 1200

Split- Phase 1 (WBL)

14.4 17.5 12.2 15 10.5 12.7 12 17 13.5 14.5 10.5 12 11 17 13 12.1 12.8 13 14 14.7 10.5

Split- Phase 2 (EBT) 24.2 27 21.7 21 22.4 44.2 23.5 57.5 23.6 30.5 35.1 22.3 17.5 18.5 29.3 23.7 31.2 20 35.2 17.9 20.5

Split- Phase 3 (NBL) 13.5 16 11 11 18.1 15.3 11 12 13.7 10.7 10.8 11.1 10.9 22 10.5 15.8 19.3 14.5 15.2 12.9 13

Split- Phase 4 (SBT) 17.9 19.5 15.1 13 19 17.8 13.5 13.5 19.2 14.3 13.6 24.6 20.6 12.5 17.2 18.4 16.7 22.5 15.6 14.5 46

Split- Phase 5 (EBL) 15.4 17.6 10.5 11 10.5 17.3 11 17.2 18.6 12 17.6 12 10.5 12 12.7 14.3 18 11 15 13.1 14

Split- Phase 6 (WBT) 23.2 26.9 23.4 25 22.4 39.6 24.5 57.3 18.5 33 28 22.3 18 23.5 29.6 21.5 26 22 34.2 19.5 17

Split- Phase 7 (SBL) 14 16 13.6 10.9 14.5 13.1 12 13 15.3 12 10.5 17.6 19 12 15.2 17.6 16 18 10.9 12.5 12.5

Split- Phase 8 (NBT) 17.4 19.5 12.5 13.1 22.6 20 12.5 12.5 17.6 13 13.9 18.1 12.5 22.5 12.5 16.6 20 19 19.9 14.9 46.5

Cycle Length

24

70 80 60 60 70 90 60 100 70 70 70 70 60 70 70 70 80 70 80 60 90

Offset W 0 34 0 33 49 48 8 64 37 42 35 2 47 29 46 63 55 28 17 31 2

Offset M 8 20 2 2 0 12 6 14 0 4 8 6 10 20 0 20 8 8 9 1 0

Offset E 44 2 10 1 8 4 58 8 55 62 6 40 38 24 39 16 10 2 18 8 34

EB Left Turn Treatment 2 2 2 2 3 2 2 3 2 3 2 2 2 2 2 2 2 2 3 3 2

WB Left Turn Treatment 2 2 2 2 3 2 2 3 2 3 2 2 2 2 2 2 2 2 3 3 2

NB Left Turn Treatment 2 3 2 3 2 2 2 2 2 2 2 2 2 3 2 3 2 2 2 2 2

SB Left Turn Treatment 2 3 2 3 2 2 2 2 2 2 2 2 2 3 2 3 2 2 2 2 2

NB Left Turn Sequence 1 1 2 1 1 2 2 1 1 2 1 2 2 1 1 1 1 2 2 1 2

SB Left Turn Sequence 2 1 1 1 2 1 2 2 2 2 2 2 1 1 2 1 2 2 1 2 2

EB Left Turn Sequence 2 1 2 2 1 1 2 1 1 1 1 2 2 1 2 1 1 1 2 2 2

1 2 1 1 1 2 1 2 2 2 2 1 2 2 2 2 2 2 2 2 1

WB Left Turn Sequence

The following key applies to the left-turn treatment and sequence columns:

Table 3. Left turn treatment PARAMETER

CODE

Left Turn Sequence Lag

1

Lead

2

Left Turn Treatment Permissive

1

Protective

2

Protected-Permissive (E/P)

2

2.2 SUMMARY In order to quantify surrogate safety measures and enhance efficiency in adaptive control, a safety performance function was developed to predict the likely changes in total conflicts (measured in SSAM) based on changes to the traffic state and choices for alternative signal timing parameters. Based on the number of traffic state inputs and traffic control variables, the possible number of combinations that could be tested is extremely large. All possible combinations cannot be tested, therefore a statistical methodology, namely a Design of Experiments (DoE) was used to identify and test meaningful combination of signal timing settings and traffic parameters. After specifying a small test network, realistic ranges for input parameters were derived and then generated randomized combinations of test cases using a “Latin Hypercube” approach. Subsequently, those scenarios were run in VISSIM to obtain efficiency performance (represented by the V/c ratio as reported by the simulation model for each phase at intersection M) and the safety performance (as represented by the total and component totals of traffic conflicts). That information was then used to train the neural network and then tested on additional scenarios to see how well the prediction works. The construction of the Safety Performance Function (SPF) using neural network is discussed in the next chapter.

25

26

3 DEVELOPMENT OF SAFETY PERFORMANCE FUNCTION As discussed in the previous section, the development of the safety performance function involved a Design of Experiments (DoE) approach to identify combinations of signal timing parameters and traffic patterns (volumes and route proportions) that would be representative of a very large potential test space. This was necessary because of the extreme time and manual effort required to evaluate each combination using simulation and post-processing the results with SSAM. In particular, the processing of trajectory data from each simulation iteration with SSAM is quite time consuming. The process adopted was an iterative one. A reasonable set of 150 test cases were generated and then used results to train the neural network. Based on the training results, another small number of cases were identified to replace some cases that resulted in oversaturated conditions and re-trained the neural network. This resulted in a reasonable prediction accuracy of 17% average error. The neural network was then embedded in the adaptive control algorithms and used for prediction of total conflicts due to anticipated changes in the signal timing parameters. This process is discussed in more detail in the following sections. 3.1 OVERVIEW OF NEURAL NETWORKS A neural network is a set of mathematical functions that relate a set of inputs to a set of outputs. Neural networks have been applied to a wide variety of prediction and pattern recognition problems in almost every area of science including process control, machine vision, and finance. The power of neural networks is particularly observed when the problem is highly nonlinear and the correlation of the inputs to the outputs is not obvious. This is certainly true in the correlation of simultaneous changes to multiple traffic signal timing parameters to the resulting total conflict rate. There are many formulations and architectures that fall under the general umbrella of neural networks including variations in the type of mathematic transform functions that are used, how many processing layers are included, and how feedback is applied (or not). In general, a neural network is modeled after the function of the brain where each neuron takes inputs (electrical signals from other neurons) and processes those signals across a synapse. When the activation level is high on the inputs the synapse “fires”, sending electrical signals to other neurons. The interconnected network of trillions of neurons in the brain builds up higher and higher level functions from this basic premise into information such as recognition of a person’s face as a friend, or the smell of apple pie and elevated temperature in the kitchen must mean that someone is baking pie in the oven. The basic mathematical concept of a neuron is illustrated in Figure 10. Input values are multiplied by weights and summed together. This sum is then input to an activation function. The most common activation function is the “sigmoid”, illustrated in Figure 11.

27

Figure 10. Illustration. Basic concept of a “neuron” in a neural network

Figure 11. Illustration. Sigmoid activation function The “network” in a neural network is the combination of the number of neurons (or nodes) and how they are arranged in layers. In our research, we have ~25 inputs (signal control parameters and traffic conditions) and one output, that is total traffic conflicts. Designing the structure of a network is itself a complicated procedure. Fortunately, there are tools that assist in optimizing network structure which is described below.

28

A neural network is “trained” to match the inputs to the outputs as closely as possible by iteratively adjusting the weights at each of the neurons. This process is called supervised learning since the correct outputs are known for the test cases. Errors that are made in predictions using the current weights are fed back into the system and adjustments are made to the weights based on a mathematical search algorithm designed to drive the error to zero. This process is illustrated in Figure 12.

Figure 12. Illustration. Illustration of the training process There are many search algorithms available to be used that have (esoteric, in many cases) different properties. The main property of concern in designing the network choosing the search method in our problem is generalization. With a small data set, such as in this research, almost any network structure can drive the prediction performance on the training set to zero. The real goal is to produce the lowest error on the set of test cases. These cases are combination of the traffic signal parameters and resulting traffic conflicts that the neural network was not trained on. In our research, we trained the network with approximately 120 cases and tested with 30 cases. Some training methods randomize the training set and the test set to improve generalization. The MATLAB software was used with the neural network toolbox to implement this procedure. MATLAB is a highly used system for mathematical calculations, originally developed for making it easy for scientists and engineers to perform matrix mathematics without writing procedural code. The neural network toolbox has a wide variety of features for neural network design, implementation and training. The principal tool used is illustrated in Figure 13. The neural network training tool allows the user to simply test the resulting performance of a wide variety of network structures, number of layers, number of neurons per layer, and training algorithms.

29

Figure 13. Illustration. Neural network training tool in MATLAB The following network structures were tested: 

Cascade forward network



Feed forward network



Fitting network

These are the most straight-forward structures. Roughly another 15 different network structures were not tested due to incompatibility with the type of input data or our assessment that the complexity introduced was not necessary for our relatively simple input-output matching problem. These three structures were tested with the baseline number of nodes and layers as suggested by MATLAB, using the conjugate-gradient back propagation training algorithm with Powell-Beal restarts (notably the first search algorithm in the list; the subsidies of the differences in the algorithms are not particularly important for this first step). At the conclusion

30

of the training process, it was found that the cascade forward structure performed best with about 25% average error on the test cases. For training methods, we then tested the following algorithms with our baseline cascade forward network: 

Conjugate gradient back propagation with Powell-Beale restarts (CGB)



Conjugate gradient back propagation with Fletcher-Reeves updates (CGF)



Conjugate gradient back propagation with Polak-Ribiére updates (CGP)



Levenberg-Marquardt (LM)



Resilient back propagation (RP)



Scaled conjugate gradient back propagation (SCG)



One-step secant back propagation (OSS)

Levenberg-Marquardt resulted in the best results on the cascade forward network driving the error a bit lower to 18%. We then tested various node and layer combinations using the cascade forward network using Levenberg-Marquardt training: 

1, 2, or 3 layers



1-20 nodes per layer

No appreciable improvement was found in the addition of more layers. In fact the performance was noticeably worse with higher numbers of nodes and layers. This is not surprising since the training data set is relatively small. The best performance was determined with one layer or five nodes at 17% average error on the test set. With ~25 inputs and 5 nodes, the number of weights to be trained (125) is roughly the same as the number of training cases. This was somewhat surprising as we originally expected fewer weights to be necessary, with perhaps two layers of processing. Adding additional layers and nodes only increases the number of weights to be determined, so as with any mathematical function fitting problem, when the number of parameters in the model exceeds the number of data points, the problem is indeterminate. Table 4 lists the performance of the neural network on the training cases. As discussed above, the performance is close to 100% correct with an average error of just 3.7%. The performance on the training set is noticeably good on larger number of total conflict rates. Table 4 lists the performance of the neural network on the test cases. As shown the relative average error is 17%. The performance is actually better than reported by average % because the results are skewed by large % errors when the numbers are very small. For example, a real total conflict rate of 40/hr and a predicted total conflict rate of 20/hr is a 50% difference (or 100%, depending on how you look at it). In reality, a 40 conflict per hour rate is very low. Overall safety is not appreciably different at 40 conflicts per hour versus 20 conflicts per hour when the reasonable ratio of conflicts to real crashes is on the order of 5,000:1 and typically reported as larger than that. 31

Table 4. Performance of neural network on training of cases

SCN 104F 94F 12F 22F 124F 96F 107D 22E 94D 124D 54D 136E 122C 62E 147F 141C 77D 99D 18C 22D 147D 129D 100D 19D 100E 105C 7D 43D 91B 46D 94B 69D 104D 62D

TOTAL CONFLICTS INTERSECTION M 920 874 816 790 750 738 734 712 687 645 627 600 574 556 510 502 486 426 421 419 419 413 412 412 404 394 376 369 361 347 335 333 326 306

PREDICTED TOTAL

ABSOLUTE DIFFERENCE

PERCENT DIFFERENCE

914 897 818 790 764 753 725 741 660 656 610 611 516 527 495 493 472 384 430 397 403 403 381 402 395 406 386 347 339 365 329 323 327 306

6 -23 -2 0 -14 -15 9 -29 27 -11 16 -11 57 29 15 10 14 41 -9 22 16 10 31 10 9 -12 -10 22 22 -18 6 10 -1 0

0.7 -2.6 -0.2 0.0 -1.9 -2.0 1.2 -4.0 4.0 -1.7 2.6 -1.8 10.0 5.2 3.0 1.9 2.9 9.7 -2.1 5.2 3.8 2.4 7.6 2.4 2.2 -3.2 -2.6 6.0 6.1 -5.2 1.8 3.0 -0.2 0.0

32

SCN 122B 31C 3C 38D 43C 99C 54C 31D 107C 38C 56C 136C 72C 77C 54B 60C 129B 128B 22C 100C 62C 118B 19C 128C 128D 1D 124C 141B 46C 34C 31B 124B 104C 105B 147B 107B 56B 54A

TOTAL CONFLICTS INTERSECTION M 290 284 278 277 275 275 274 273 272 256 249 247 247 244 241 235 234 225 223 221 218 215 214 212 212 210 209 205 190 185 184 173 172 170 167 165 157 155

PREDICTED TOTAL

ABSOLUTE DIFFERENCE

PERCENT DIFFERENCE

302 262 281 288 280 262 304 268 312 273 241 266 214 250 250 241 194 164 220 207 202 186 227 196 187 193 198 174 180 185 185 182 195 173 172 181 145 113

-12 22 -2 -11 -4 12 -30 4 -40 -17 8 -19 33 -6 -9 -6 40 61 3 14 16 29 -12 16 25 17 10 31 10 -1 -1 -9 -23 -3 -5 -16 12 42

-4.3 7.8 -0.9 -4.1 -1.6 4.5 -10.9 1.6 -14.6 -6.7 3.0 -7.7 13.4 -2.5 -3.7 -2.5 17.0 27.2 1.3 6.1 7.5 13.4 -5.8 7.3 11.9 8.0 5.0 15.0 5.5 -0.3 -0.4 -5.4 -13.5 -1.8 -2.9 -9.8 7.4 26.9

33

SCN 19B 69C 77B 46B 96B 22B 100B 16C 1B 1C 94A 43B 38B 147A 107A 34B 60B 104B 38A 46A 56A 25A 122A 141A 55B 34A 16B 118A 16A 31A 105A 124A 69B 43A 104A 55A 96A 100A

TOTAL CONFLICTS INTERSECTION M 152 151 146 143 142 141 140 126 125 123 122 121 118 112 110 98 98 93 91 90 88 83 76 75 75 74 71 64 63 63 62 62 59 58 52 46 45 41

PREDICTED TOTAL

ABSOLUTE DIFFERENCE

PERCENT DIFFERENCE

129 152 132 148 121 120 187 124 113 152 94 135 83 95 141 92 88 64 63 78 84 61 80 73 29 30 71 42 85 58 81 70 51 38 29 23 48 50

23 -1 15 -5 22 21 -48 1 12 -29 28 -15 35 17 -31 6 10 30 28 12 4 22 -4 2 46 44 0 22 -23 4 -19 -9 8 20 23 23 -3 -9

15.1 -0.7 10.0 -3.7 15.1 15.2 -34.1 1.1 9.4 -23.8 23.1 -12.3 29.5 15.5 -28.0 6.3 9.8 31.8 30.5 13.7 4.4 26.1 -4.7 3.1 61.1 59.1 -0.3 34.0 -36.0 7.2 -30.5 -14.2 14.1 34.6 44.0 49.2 -7.2 -21.7

34

TOTAL CONFLICTS SCN INTERSECTION M 129A 38 22A 35 77A 32 60A 31 19A 28 1A 12 128A 11 69A 9

PREDICTED TOTAL

ABSOLUTE DIFFERENCE

PERCENT DIFFERENCE

65 61 54 18 27 10 57 58

-27 -25 -22 13 1 3 -46 -48

-70.9 -72.0 -67.8 42.6 3.9 20.7 -416.9 -514.4

6.203291736

-3.71630062

35

3.2 SUMMARY A safety performance function (SPF) was developed by training a cascade feed-forward neural network to learn the relationship between the signal timing settings, efficiency measures, and the resulting average traffic conflict rates. The average traffic conflict rates were post-processed using SSAM from trajectory data from the VISSIM simulation model. The Virtual D4 controller was used to implement the signal operation in VISSIM. From training the network with approximate 150 scenarios, the average error on the prediction of traffic conflicts, on cases that were not used for training, was 17%. This average error was considered as an acceptable performance to use in the network as the safety performance prediction function in the adaptive control system. MATLAB was used to implement the neural network and export the network calculations as a C++ DLL that was integrated with the rest of the adaptive system code. In the next section, we discuss the adaptive algorithms and describe how the safety performance prediction function is used to balance efficiency and safety. Following the description of the adaptive algorithms, we present a series of simulation scenarios which were constructed to demonstrate and verify the basic capabilities of the real-time adaptive efficiency and safety algorithms. The safety performance function (SPF) is not an optimizer like the efficiency Model, but rather a performance index of surrogate safety measures that is derived based on the selected signal timing settings. When the SPF is enabled, the new signal timing settings will be implemented in the controller only if the safety performance function derives a lesser number of conflicts than with the existing signal timing plans. Otherwise, the current signal timing settings will be retained. If the SPF is disabled, then all new optimized signal timing settings will be implemented in the controller.

36

4 DEVELOPMENT OF MULTI-OBJECTIVE ALGORITHMS 4.1 OVERVIEW This project has focused on developing algorithms that can optimize traffic signal timing to balance the performance benefits for safety and efficiency. The safety performance function discussed in earlier chapters provides the tool that estimates the safety effect of modifications to signal timing parameters by predicting the number of conflicts that will result from the modified signal timing. By comparing the predicted number of conflicts for the current settings and potential settings, each of the adaptive algorithms can balance the needs to improve safety, if enabled, with efficiency by rejecting changes that improve efficiency but degrade safety. It was determined based on casual discussions with practicing municipal traffic engineers that it would not be politically acceptable to accept efficiency improvements that degrade safety by increasing the risk of crashes based on a prediction of an increase in the number of conflicts. Each of the adaptive adjustment algorithms thus solve first for new signal timing settings that improve efficiency and then check the prediction of the impact of those settings on safety. If the number of conflicts increases, the adjustment is abandoned and the current timing settings are retained. If the number of conflicts is reduced or statistically similar to the current prediction of total conflicts, the modifications are implemented and downloaded to the field controllers. As will be shown in the test results, this makes the adaptive system, in general, make fewer adjustments to the signal settings. In most test cases the number of aborted adjustments is reasonably low, which is encouraging since this indicates that modifying signal timings in an adaptive manner typically improves both safety and efficiency together, or the adjustments that improve efficiency have no marginal detrimental effect on safety as measured by the estimated total conflict rate which is overwhelmingly driven by the rear end conflict rate. By smoothing traffic flow and eliminating as many unnecessary stops, efficiency and safety are both improved. Although initially started with the idea of developing five tuning algorithms, the final real-time adaptive methodology comprised of four principle algorithms for tuning signal splits, offsets, cycle time, and phase sequence. Switching between protected, protected-permitted, and permitted-only left turns was considered based on the significance of the effects on safety, but was not implemented. Discussion with practicing municipal engineers indicated that while the concept was promising, there are significant challenges in implementation of such a methodology related to both field infrastructure (i.e. need to install flashing yellow arrow or fivesection heads) and driver perception. Informal tests of modifications of protected-permitted and permitted only displays by time-of-day by some of the engineers contacted indicated that driver confusion was significant. The adaptive system builds upon the operation of the ACS Lite version concept developed between 2002-2007 for FHWA and industry consortium by Dr. Gettman (currently with KHA), Dr. Shelby (currently with Econolite), and Dr. Head,(currently with University of Arizona (Gettman, et., al. 2006). Under this project, the new developed adaptive methodology is named Kadence. In the approach to the ASC Lite version, new signal timing parameters are downloaded to field controllers every 3-4 cycles. The field controller then begins operating in an actuated-coordinated or actuated-free mode with these new settings. Based on past experience with traffic control systems that override the controller’s timings every second (RHODES,

37

OPAC, SCOOT, SCATS, and InSync), this methodology of adjusting timings is more effective and less error prone. In addition, the approach of the ACS Lite version requires minimal capital investment, infrastructure, detectors, configuration, and calibration since existing detection, communications, and field controller hardware are supported. The system operation with the lite version of ACS has been validated in over 10 deployments nationwide to produce improvements to travel time and system delay over actuated-coordinated operation with TOD plans. In addition to the implementation of three optimization algorithms (cycle tuning, cycle selection, and phase sequence selection), several other improvements were implemented. Kadence requires three cycles of good observations of phase timing and detector data before making decisions but now checks every 30 seconds to identify if that requirement is satisfied. This vastly improves the responsiveness of the system versus the fixed-horizon scheme (5 minutes, 10 minutes) implemented in the FHWA ACS system v1.4.2 circa May 2007. The offset tuning algorithm was enhanced to search offsets in a range instead of only considering fixed changes (+5, 0, -5). By selecting larger search bounds (say, +/-20s), Kadence can quickly find the correct offset solution when the current solution is particularly poor. Kadence is already integrated with VISSIM using Virtual D4 controller firmware. This provided the capability to rapidly prototype and test real-world situations with accurate controller parameters as detailed in further chapters. For field deployment, the approach is based upon second-by-second polling of field controllers with standard protocols and data formats (NTCIP, AB3418E) so it is able to be implemented with the widest possible variety of field infrastructure. 4.2 FIELD DETECTOR REQUIREMENTS Detection requirements for the adaptive system are illustrated in Figure 14 and discussed as follows: 

Stop-bar detection for all phases that are adaptively controlled. Stop-bar detection is used for split adjustment, cycle adjustment and phase sequence adjustment. These detectors can be any length. Lane-by-lane detection (i.e. separate lead-in cables from the field loops to the cabinet) is strongly preferred, but not required. Any detection technology (video, loops, magnetometer) is supported that provides contact-closure inputs to the traffic controller. Any phase can be designated to be excluded from optimization, if desired, or if the phase is anticipated to be run in recall because of lack of detection.



Advance detection for all coordinated phases for tuning offsets and making phase sequence adjustments. These loops or zones should be relatively short (i.e. 6x6 zones or loops in each lane). Lane-by-lane detection (i.e. separate lead-in cables from the field loops to the cabinet) is strongly preferred, but not required. Any detection technology (video, loops, magnetometer, radar) is supported. Exit detection is supported, as is midblock detection. Re-use of detection zones for phase extension (e.g. 150’-400’) upstream is the most economical way to achieve the necessary operation. Installation of additional mid-block or exit zones is not necessary.

38

Figure 14. Illustration. Stop bar and advance detection Kadence was developed for the Windows PC platform and requires no additional field hardware in field traffic cabinet. The system can be deployed as part of a centralized/distributed signal control system or as a stand-alone management platform like a “master” controller using a fieldhardened PC with solid-state HD and fan-less processor and associated type 2070 and NEMArated peripherals. 4.3 ADAPTIVE ALGORITHMS The lite version of ACS optimization system was developed by Dr. Gettman (currently with KHA), Dr. Shelby (currently with Econolite), and Dr. Head (currently with University of Arizona) for FHWA from 2002-2007 (Gettman, et. al. 2006). Field trials and deployment of the lite version of ACS sites was completed in 2006. All field deployments resulted in improvements to system delay and arterial travel time. The lite version of ACS was developed to address the financial costs of the high sensor density and communications infrastructure required by the adaptive algorithms developed in the mid-1990s. This project for USDOT revisited those methods and expanded upon the capabilities of the system. Kadence is thus based on a simple traffic model that is self-calibrating. The algorithms tune the signal timing parameters according to simple predictions in the changes to the measures directly and rules for adjusting parameters based on the collective measured values. The lite version of ACS originally consisted of only split and offset tuning methods. Kadence includes additional algorithms for cycle selection, cycle tuning, and phase sequence selection as well as enhanced offset tuning algorithm. All of the algorithms include the consideration of safety performance in the same manner. The adjustments are first computed to improve efficiency according to the algorithms specified in the following sections. If those adjustments are detrimental to safety as predicted by the safety performance function, the current signal

39

settings are retained and that adjustment is not implemented. The system operates using a baseline TOD schedule of coordinated timing plans. The fixed values of the cycle, splits, offsets, and phase sequence are implemented at the beginning of the period according to the TOD schedule. Kadence then considers adaptively changing the values according to the algorithms described in the following sections. If the TOD schedule includes a period where one or more signals are operated in “free” mode, currently Kadence does not tune any parameters. 4.3.1 Cycle Tuning Cycle time is adjusted on a section- or arterial-wide basis to provide adequate capacity to operate all of the signals under capacity. Kadence uses a heuristic rule to adjust the cycle time up or down a given step size. In a straight-forward fashion, if the cycle time is increased by 4 seconds, then every phase on the controller gets a proportion of the additional time. For example, if there are four phases per ring, one additional second is provided for each phase split. The split adjustment algorithm, operating under a different optimization criteria than the cycle length algorithm, will refine the splits at a later step if this allocation results in uneven phase utilization. The step size is user-defined. Minimum and Maximum cycle limitations are imposed including limitations by minimum green, pedestrian clearance times, and user-defined minimum and maximum cycles. To avoid instability from changing the cycle length frequently, there must be at least 3 cycles of vehicle-occupancy data for critical phase utilization monitoring detectors in the system to execute the cycle tuning algorithm. This methodology will favor longer cycles during peak periods as traffic demand builds, which is generally accepted as an appropriate strategy. However, recent research (NCHRP 03-90) is indicating that when the conditions are extremely oversaturated, shorter cycles will provide more efficient throughput (Gettman, et. al., 2012). These improvements or algorithms have not yet been integrated into Kadence but are planned for future work. This will improve the capability of Kadence to provide sound decisions during incident response conditions, such as heavy diversion of flows from a freeway to a parallel arterial or frontage road system. The cycle tuning algorithm used in Kadence is illustrated in Figure 15; it extends from a “critical intersection” algorithm to consider saturation levels at multiple intersections in the adaptive group. The phases that are designated to be checked in the cycle tuning algorithm are determined by the user. When the average phase utilization on these critical phases is above the user-defined threshold (say, 80% phase utilization) to increase the cycle, a given (the userdefined step size) number of seconds are added to the cycle time. Similarly, the system cycle time is decreased by a fixed number of seconds when the average of the phase utilization on the critical phases is less than a lower threshold (say, 50% phase utilization). If either of those changes results in a predicted increase in the number of conflicts for the system, the current cycle time is retained as long as the “enable Safety Evaluation” function is enabled.

40

Figure 15. Illustration. Flow chart of cycle time tuning algorithm In this manner, Kadence considers the possible improvements to both safety and efficiency by first checking the efficiency metric for the cycle time – is the average saturation level of the critical phases above or below the designated thresholds? If the answer is yes then an increase or decrease to the cycle time is executed. If this change results in improvement to the safety metric (i.e. reducing the total predicted conflicts), Kadence adjusts the cycle higher or lower. If the adjustment results in a predicted degradation to the total conflicts, the current cycle time is retained. This algorithm provides predictable and controllable changes to cycle time. 4.3.2 Cycle Selection In addition to cycle tuning, an algorithm has been developed, which evaluates a decision to either implement the cycle time that is next in the TOD schedule earlier or later than was originally planned. This “Enable Plan Switching Adjustment” function can be used in conjunction with the cycle tuning method or alone. This allows the system to adapt to changes to the beginning or end of peak periods. The same evaluation approach for the incremental cycle tuning algorithm 41

described previously is used but rather than changing the cycle just a few seconds, the system enables the new cycle time immediately and recalculates splits and offsets appropriately. This is illustrated in Figure 16. The “Enable Incremental Cycle Adjustment” function can be enabled concurrently with the cycle tuning and “Plan Switching Adjustment” function.

Figure 16. Illustration. Concept of implementing the next cycle time earlier or later than scheduled As illustrated in Figure 17, the algorithm begins considering implementing the next pattern early in the TOD schedule if it meets a phase utilization threshold for a configurable number of minutes before the next pattern change time. For example, this threshold time might be set to 30 minutes prior to the schedule change so if the pattern is scheduled to be adjusted at 10:00 am, the algorithm will begin considering implementing the next pattern at 9:30 am. If the thresholds for phase utilization are not exceeded to implement a lower or higher cycle, the current cycle is retained. If the phase utilization thresholds are not exceeded after the scheduled time to change to the next cycle, the system can keep the current pattern in operation. After a configurable amount of time, however, the system will transition to the next pattern in the TOD schedule. For example, if this threshold time is set to 30 minutes and the pattern is scheduled to be adjusted at 10:00 am, the system may retain the AM peak pattern until 10:30 if the saturation levels indicate that it is more efficient to do so. At 10:30, the system will then implement the mid-day pattern.

42

Figure 17. Illustration. Cycle selection process 4.3.3 Offset Tuning Cycle time tuning affects all of the intersections in the network if they are all operating in a coordinated mode. Tuning offsets improves progression performance along primary routes for phases that are coordinated. Offset tuning algorithms are particularly straightforward. The proven and robust methodology used in the lite version of ACS is also used in Kadence with several key enhancements as described below, in addition to the consideration of the balance between efficiency and safety.

43

The concept of the data-driven offset adjustment algorithm is summarized in two simple statements: (1) use detectors at least several hundred feet upstream of the signal to construct cycle-based (“cyclic”) profiles of traffic flow arriving to the intersection, as shown in Figure 18 and (2) adjust the offset to maximize the number of vehicles arriving during the green phase. Periodically, small, incremental adjustments are made to the offset to maximize the total proportion of cyclic flow arriving to a green light. This concept is then expanded to consider and mitigate the effects of such modifications to the offset value for multiple approaches (including the consideration of cross-coordination on all four approaches) and the effects of changes at a given intersection on adjacent intersections. The algorithm begins from the offset value of the TOD pattern. A user-configurable maximum deviation from the original setting (either an increase or decrease to the offset value) is defined for each offset to restrict the algorithm (if desired) from drifting too far away from the original solution. The user can also specify that this value is unbounded, which allows the system to search for any offset. For example, if the initial offset is 20s and the maximum deviation is set to 10s, the algorithm will be restricted to implement offsets with the range of 10s to 30s. Figure 18 illustrates the detector locations used for offset tuning. There is one detector station for each coordinated approach. Intersection 1 is referred to as the upstream intersection and intersection 2 is referred to as the downstream intersection. Traffic volume and occupancy is measured at some point between 1 and 2 by a detector in each lane. These detectors can be located where typical advanced detectors are located (150-400 feet from the intersection). Placing detectors further upstream can improve the quality of the flow rate measurements, and reduces the possibility of vehicles queuing over the detectors when the light is red. It is not necessary to have one detector per lane returned to the controller in a separate amplifier, but this practice will improve performance of the algorithm. Exit detection or mid-block detection can also be used, if available.

1

2

Flow Profile Detectors Figure 18. Illustration. Typical flow profile detector locations on coordinated approaches Assuming that the traffic signals at both intersections are using the same cycle length, and that traffic volumes and turning proportions are reasonably steady over time, it is expected that the

44

detector will measure approximately similar recurring patterns of flow each cycle. These patterns of flow, as shown in Figure 19, are referred to as cyclic flow profiles. Plots of the flow profile data (traffic volume count and occupancy observations) as a function of the local cycle time of the controller (time is on the x-axis; not direction from West to East) are shown in Figure 19. The magnitude of the occupancy is indicated by the height of their corresponding bars in each row of the chart. The height of the bars in each row is scaled by the maximum value observed in that row (equal heights in different rows do not necessarily indicate the same occupancy value). These profiles indicate that it is typical over the last few cycles that traffic is arriving near the beginning of the local cycle time for this approach. Secondary platoons and individual vehicles also show up throughout the cycle, due to turning flow on the cross street phases. Occupancy Volume Occupancy

Cycle 6

Cycle 5

Volume Occupancy Volume

Cycle 4

Occupancy Volume

Cycle 3

Occupancy Volume

Cycle 2

Occupancy Cycle 1 Volume

Local Cycle Time

Figure 19. Illustration. Example of traffic volume count and occupancy data from a typical advance detector Figure 20 illustrates an example of phase timing history observed over the last several cycles at an intersection.

45

Figure 20. Illustration. Example of phase timing for each of the last several cycles The number and color of each column in the timeline corresponds to the active phase interval (green, yellow, and red) displayed by each ring at that time in the cycle. All subsequent cycles shown below the first row are actual data recorded from a field test controller with the most recent data at the top and progressing back in time as you go further down the display. Each cycle begins at the local zero time, which is labeled on the left. Note that the duration of non-sync, actuated phases (typically phases other than 2 and 6) are variable, and one or more of phases may be skipped in any given cycle according to the actuated operation of the traffic controller. Thus, the time at which the controller returns to the sync phases (typically phases 2 and 6) can and does vary from cycle to cycle. The cyclic flow profiles are then averaged to generate a single, representative cyclic flow profile for the flow profile detector. Each link, and thus each flow profile detector on that link, is associated (via user configuration) with a corresponding phase at the downstream intersection. Note that in the case of an arterial, the progression phases generally correspond to the coordinated phases on the controller, but any phase can be designated as the progression phase if the primary progression movement is a turning flow. In a grid network all major through phases (coordinated and non-coordinated) might be configured as progression phases for their corresponding approaches. Figure 21 shows that during a portion of the cycle, the progression phase is green 100% of the time, starting at local time 50 and ending at local cycle time 0 (or 80). Each cycle one or more non-sync phases typically gap out early, and in such cases, the controller returns to the coordinated phase earlier than is required (and is typically termed “early return to 46

green”). Figure 21 illustrates this early return to green behavior with the tapering percent-green bars prior (to the left of) to the programmed start of main street green split (local time 50 seconds). As shown, this progression phase started as early as local cycle time 27 in at least one cycle during the last ~10 cycles.

Volume Occupancy Phase Green Probability

Local Cycle Time Figure 21. Illustration. Example of average cyclic volume and occupancy profiles Note that occupancy, rather than volume, is the preferred detector measurement used to generate flow profiles. The flow profile scenario shown above is an example of the performance of a good offset for one-directional travel. The arriving platoons are indicated by the cluster of relatively tall occupancy bars between local cycle time 40 and local cycle time 75 which corresponds to the green of the service phase. 4.3.3.1 Captured Flow The effectiveness of offset settings at upstream and downstream intersections is measured or quantified by calculating the % arrivals on green, or the “captured flow”. This percent arrivals on green measure is a surrogate measure for stops and delay. Specifically, the captured or progressed flow is the amount of flow (in units of vehicle-seconds of occupancy) arriving to the stop-line at a given point in the cycle multiplied by the percent of time the progression phase is green at that time during the cycle. The algorithm evaluates different offsets by calculating the captured flow on each approach and selecting the offset that maximizes the total amount of captured flow. Distributed Offset Adjustment: Kadence uses a distributed offset adjustment method. This approach makes offset adjustments for each controller with consideration of the effects of each decision on adjacent signals. Each controller considers a range of offset settings: no change, adjust up to  seconds earlier, or adjust up to  seconds later. The adjustment maximum step size,  , is a user-configurable value. If the value is set at 10, for example, Kadence will search offsets in each step in the range of (+10, +9, +8, +7, …., 0, -1, -2, -3,….., -10). The adjustment procedure estimates the amount of cyclic traffic flow progressed for inbound and outbound links of the controller for each of the adjustment options and chooses the option that maximizes the total progressed flow. As a stability measure, there must be at least 3 cycles of vehicleoccupancy data for every flow profile detector at this intersection and adjacent intersections to execute the offset tuning algorithm. This provides a level of smoothing so that the changes are 47

not frequently toggling between one offset and another. The lite version of ACS project field tests determined that this methodology is effective at making adjustments to offsets to improve progression, while mitigating the effects of controller transition by only making small adjustments for each tuning. If the difference in the % arrivals on green between the evaluated offsets is not greater than small amount of improvement, say, 5%, the controller will remain at the current offset. This reduces transition events that do not result in significant improvement to performance. Including Safety in the Assessment of Offset Performance: Similar to other algorithms, the approach to tuning offsets, as discussed above, is augmented to consider both efficiency and safety. Similar to the approach for tuning the cycle time, the algorithm uses the safety performance function as the performance calculator. This approach is detailed in Figure 22In the Phase 1 research, it was concluded that adjustments to offsets had the weakest correlation to improvements or detriments in safety.

Figure 22. Illustration. Offset adjustment algorithm flow chart

48

As shown above the approach is to identify the offset that best improves the % arrivals on green. If the best alternative offset is greater than a 5% improvement over the current offset’s % arrivals on green, then the safety prediction is considered. If efficiencies decrease the total number of conflicts, the offset will be adjusted. If the change is predicted to increase the total number of conflicts, the algorithm will not adjust the offset. 4.3.4 Split Tuning Splits are tuned by collecting volume and occupancy data from detectors at the stop bar of the intersection in the same way that was implemented in the lite version of ACS project which is similar to the methods used by SCATS and SCOOT. The algorithm attempts to equalize the degree of saturation on all the phases at the intersection. To be specific, the algorithm in Kadence minimizes the maximum degree of saturation on any phase, rather than driving all of the saturation levels to the same value. This algorithm also allows coordinated phases (or any phase, but this biasing is typically applied to coordinated phases) to have biased splits, so that progression is protected when the saturation level of the coordinated phase is lower than that of side-streets. Without such biasing, split adjustment methods that equalize the degree of saturation on all phases tend to focus more on providing adequate LOS on side streets while degrading progression along a critical route. Any phase can also be determined to be left out from the split tuning process. This is commonly applied to phases on minor side-streets that very occasionally experience a surging traffic flows that do not last more than a few minutes. In the absence of regular arrivals, the splits will be adjusted to the minimum possible value. Since the absence of traffic on the side-street will naturally result in additional time to the main street when the phase is skipped, the side-street phase may be set by the traffic engineer to a reasonable value that provides adequate LOS during the burst and kept fixed. Kadence will adjust the other parameters as appropriate. A good example for this would be the exit from a church, a theater, or a convention center. 4.3.4.1 Split Constraints There are constraints on split adjustments which can be defined using a phase-barrier diagram. A typical phase-barrier sequence is illustrated, with barriers explicitly labeled in Figure 23. b

Ø1

Ø2

Ø5

Ø6

A

Ø3

Ø4

Ø7

Ø8

b

Figure 23. Illustration. Ring diagram with barriers In discussing the split adjustment algorithms, it is necessary to refer to certain groups of phases. The collection of phases on a particular ring, between two particular barriers is referred to as a ring group. There are four ring groups, (1, 2), (3, 4), (5, 6), and (7, 8). A barrier group, is the collection of all phases (or all ring groups) between two particular barriers. In the example above, there are two barrier groups (1, 2, 5, 6) and (3, 4, 7, 8). The algorithm uses these groups to swap split time from one phase to another in order to determine the set of split adjustments that can be made that still satisfy the cycle time.

49

It is necessary to first determine the range of adjustments for each split, before solving for the set of split values recommended for a given controller. This includes consideration of minimum times, maximum times and pedestrian crossing restrictions. These constraints are also important to calculate for each ring and barrier group. While this may seem trivial, it is important to accurately calculate these values before searching the optimization space as this significantly improves the calculation efficiency. 4.3.4.2 Calculation of Duration Constraints 1. Compute the minimum, current, and (initial) maximum split durations for each phase p, p∈P. max initial clear a. Compute s min = max{g min ,g walk + g ped } + y p + rp , for each phase p. p p ,g p p p Note that the ASC MIB objects uploaded from each controller are in mixed precision (some in seconds, others in tenths of a second). These values are combined in tenths-of-a-second units, and the rounded up to the nearest second. If phase p is omitted in the current pattern, then s min is set to zero. p

b. The value s cur is an ASC MIB object uploaded directly from the controller for p each pattern. If phase p is omitted in the current pattern, then s cur is set to zero. p c. Initially set s max is set depending on the maximum = g max + y p + rp , where g max p p p mode (max1, max2, or max inhibit) currently used by the controller. If max inhibit is the current mode, then s max is set to 255 seconds. If phase p is omitted p in the current pattern, then s max is set to zero. Note that g max is specified in p p seconds, whereas yellow and red intervals can be specified in tenths-of-a-second. These values are combined in tenths-of-a-second precision and then rounded down to the nearest second such that s max is in seconds. p d. Ensure the current split is within the configured minimum and maximum max ≤ scur constraints. If it is not true that smin p p ≤ s p , then STOP; the configuration data is invalid. e. If the adjustment process is constrained such that only incremental adjustments from the current value are allowed, or a maximum cumulative deviation from an underlying “baseline” split value, then adjust the minimum and maximum constraints here. If the constraint on the cumulative deviation from the baseline split is not satisfied by the current split value, then STOP; this constraint is not currently achievable. f. If there are no detectors associated with the current phase, then STOP; this is a configuration error.

50

2. For each group, compute the minimum, current, and (initial) maximum durations for each group g. a. For each ring r, compute the minimum, current, and (initial) maximum ring-group duration for the ring-group in barrier group g on ring r as follows:

drgmin = ∑

min p∈Prg p

s

cur cur max max , drg = ∑p∈P sp , and drg = ∑p∈P sp . rg

rg

{ }

b. Now compute the barrier group durations as follows: dgmin = max r∈R drgmin , d

cur g

{ }, d

= max r∈R d

cur rg

max g

{

= min r ∈{R:d max > 0} d rg

max rg

}.

In general, there will always be

at least one phase in each barrier group, hence there will be one ring r such that d rgmax > 0 ; but if not, then set d gmax = 0. 3. Calculate the revised maximum duration for each barrier group, its corresponding ringgroups, and the maximum phase splits of their corresponding phases. a. Calculate the revised maximum duration for group g, under the assumption that each other group must time at least its minimum duration, as follows:

dgmax = C − ∑

min g ′∈{G:g ′ ≠g } g ′

d

.

b. Revise the maximum duration for each ring-group on each ring r of the barrier group g, such that it is not greater than the maximum barrier group duration of group g, as follows: d rgmax = min {d gmax , d rgmax }. c. Ensure that the current ring-group duration (i.e., sum of splits) is not greater than the maximum ring-group duration possible with the cycle. If drgcur > drgmax , then STOP, there is a configuration problem. d. Calculate the revised maximum splits for each phase p in the ring-groups from each ring r of the barrier group g, under the assumption that the maximum duration of the ring-group may not be exceeded and each other phase in the ringgroup must time at least its minimum duration. This is calculated as follows: smax = drgmax − ∑p ∈ P :p ≠ p smin . Ensure that the current split is not greater than the p ′ { rg ′} p ′ max revised maximum split. If scur p > s p , then STOP, there is a configuration problem. Once the minimum and maximum phase split duration constraints have been determined using the preceding procedure, the multi-ring split adjustment algorithm may commence. 4.3.4.3 Estimating Phase Utilization The split adjustment algorithm is based on the notion of balancing the utilization of all phases of a signal controller. In this section we discuss:

51



What phase utilization is,



How it is measured, and



How it is estimated for split durations other than what is currently in use by the controller.

Phase utilization is the effectively utilized percentage of available split time. It is analogous to the degree of saturation concept, which is also referred to as the volume-to-capacity or V/C ratio. Utilization is used instead of degree of saturation since the degree of saturation is calculated using volume and capacity counts or rates. Utilization is calculated using ratios of used green time to available green time. The used green time comes from the occupancy of the detector. Figure 24 illustrates a typical detector layout for measuring phase utilization with a detector placed at the stop bar for each lane. Each detector is associated with the phase that serves traffic flowing through its corresponding lane. Detector dimensions do not have to be the theoretically “best possible length”.

Figure 24. Illustration. Detector layout The methodology measures the demand for a phase by monitoring the occupancy of the phase during green. Demand is measured in terms of time, rather than volume. Utilization is given by the ratio of demand time to available green time. This is illustrated in Figure 25.

52

70% (a) Fixed-Time Control Detector Presence Phase Green Time

7 sec 10 sec (b) Coordinated-Actuated Control

58%

Detector Presence Phase Green Time 2 sec 8 sec 2 sec actual green time (10 sec) split green time (10 sec) available green time (12 sec)

Figure 25. Illustration. Measuring phase utilization for coordinated-actuated controllers Kadence polls controllers for phase timing and detector data and aggregates the data over time to construct estimates of occupancy during green, green phase duration, and utilization by cycle. The example above shows where ten seconds of green is served, seven seconds of which the detector is occupied. For a fixed-time controller, this corresponds to 70% utilization. However, in the context of coordinated, actuated controller the capacity of the phase is measured as the amount of available green time. In this case, the phase started timing green 2 seconds early due to a prior phase gapping out early. In this case, it could serve up to 12 seconds of green until it is forced-off, but it gaps out after 10 seconds of green. The notion of available green includes the remaining time to the force-off or maximum green, whichever comes first. It is also important to consider when phases have been skipped, due to lack of demand. Similarly to the other adjustment algorithms, the splits will not be tuned without at least collection of at least three observations of each phase to be tuned, while the controller is in coordination (i.e. when the controller is in preemption, the algorithm will not use this data in tuning splits). Having satisfied this condition, the algorithm then calculates the utilization of alternate split durations for each phase using the following variables and equations in Table 5.

53

Table 5. Variables used in split utilization

Np

= Number of observations of phase p

oip

= Occupancy of phase p green time during observation i (0-100%)

g ip

= Green time served by phase p during observation i (seconds)

aip

= Available green time for phase p during observation i (seconds)

dp

= Average demand (seconds occupied green) per cycle for phase p

ap

= Average available green time (seconds) per cycle for phase p

cp

= Clearance time (seconds of yellow and all-red) associated with phase p

u pt

= Utilization estimate for phase p with split of duration t seconds

4.3.4.4 Estimation of Split Utilization 1. If N p  3 , then STOP. There is not enough data for split adjustment.

2. Compute average demand, d p

 Np   o i gp i  i 1  Np

    .

p

 Np    aip   i 1  . 3. Compute average available green time, a p   Np ax 4. For each feasible split duration t, between s min and sm , estimate the utilization of that split p p assuming current demand remains the same and the phase starts on average at the same point in the cycle.

a. If

(i.e. the available green time of such a split duration is non-

zero), then set up t

dp ap  t  scp

.u

 

54

r

b. Otherwise if

(i.e. there is no available green time), then either set

u p t  0 if d p  0 (i.e. there is no demand), or set up t if d p  0. After calculating these estimated utilization levels for all alternate split durations, an iterative algorithm is executed to select the combination of split values that satisfies the constraints on  minimizes the maximum utilization each phase and phase group (i.e. ring and barrier groups) and of any phase on the controller. This procedure is discussed in the next section. 4.3.4.5 Balancing Utilization Levels Each phase is assigned a utilization measure that approximates the degree of saturation of that phase, which ranges from 0 to 100% (or higher if oversaturated). Utilization is estimated for each phase, for the full range of possible split durations, as discussed in preceding sections. Figure 26 is an example of utilization estimates for a dual-ring, eight-phase controller, where the utilization of phase 3 is very high. Table 6. of the estimated utilization of phases after the algorithm has adjusted the splits to minimize the maximum utilization of any phase on the controller. The yellow cells highlighted show the change of the most utilized phase from 100% to 72%. Also note that the new maximum phase utilization is now on phase 6, which has increased from 72% utilization to an estimated 76% utilization. Note that the result of the adjustment is not the exact same utilization level on each phase. (a) Pre-Adjustment Settings

(b) Post-Adjustment Settings

100%

100%

72%

90%

100%

80%

80%

70%

70%

Utilization

Utilization

90%

60% 50% 40%

60% 50% 40%

30%

30%

20%

20%

10%

10%

0%

0% 1

2

3

4

5

6

7

8

1

Phase

2

3

4

5

6

7

8

Phase

Figure 26. Illustration. Utilization of phases before and after split adjustment

55

Table 6. Example utilization of phases before and after split adjustment Split Times (sec) PreAdjustment PostAdjustment Utilization (%) PreAdjustment PostAdjustment

Phase 1

Phase 2

Phase 3

Phase 4

Phase 5

Phase 6

Phase 7

Phase 8

14

21

13

22

10

25

15

20

11

21

16

22

10

22

17

21

Phase 1

Phase 2

Phase 3

Phase 4

Phase 5

Phase 6

Phase 7

Phase 8

Maximum Utilization

36%

61%

100%

72%

35%

72%

40%

34%

100%

53%

61%

72%

72%

35%

76%

33%

32%

76%

As indicated previously, the primary objective of the split adjustment algorithm is to minimize the maximum degree of utilization across all phases. These objectives achieve the general effect of balancing the degree of utilization across phases and will have the effect of minimizing delay. 4.3.4.6 Incorporating Safety Assessment in the Optimization of Splits Similar to the approach used for cycle time and offset tuning, the algorithm incorporates evaluation of safety into the optimization by utilizing the safety performance function. If the total conflicts are predicted to be increased based on the reallocation of the splits, the current splits are retained. This process is illustrated in Figure 27.

56

Figure 27. Illustration. Flow chart of the split optimization process including safety analysis 4.4 PHASE SEQUENCE CHANGES Phase sequence affects both progression and delay at an intersection with respect to measures of efficiency. Under the Phase 1 research, we found that the sequence of left turn phasing can also affect safety of the intersection, particularly when an intersection operates in coordination. In our parameter tuning approach, improvements to both safety and efficiency can be made by analyzing the phase utilization measures for each signal phase and the total capture efficiency of the coordinated phases. To illustrate the concept of the algorithm, one barrier group (e.g. phases 1,2,5,6 in a standard dual-ring quad intersection) is considered. The equivalent rules apply to phases (3,4,7 and8) in the other barrier group. In lieu of a flow chart the concept is presented as a table of decision rules, as illustrated in Table 7. In the left column, the current operating phase sequence is listed, lead-lead, lead-lag, lag-lead, and lag-lag. The top row lists the potential phase sequence that we will evaluate changing to. The cells of Table 7 indicate the rules that would justify a change from one sequence to another.

57

Table 7. Rules to evaluate to consider changing phase sequence POTENTIAL NEXT SEQUENCE Lead-lead (12|56) N/A Lead-lead (12|56)

CURRENT SEQUENCE

PU5 > heavy Lead-lag (12|65)

Lag-lead (21|56) PU1 < light

Diagonal(16) dominates back diagonal(25)

Back PU5 < light diagonal(25) Lag(26) dominates diagonal(16) dominates lead(15)

N/A

PU1 < light

Diagonal(15) dominates back diagonal(26)

PU5 heavy Lag(25) dominates lead(16)

PU1 > heavy Lag-lead (21|56)

Lead-lag (12|65) PU5 < light

PU1 > heavy N/A

Back PU5 < light diagonal(15) dominates Lag(16) diagonal(26) dominates lead(25)

Lag-lag (21|65) PU1 < light

PU1 < light

> Back diagonal(26) dominates diagonal(15)

PU5 < light Diagonal(26) dominates back diagonal(15)

> N/A PU1 > heavy PU5 heavy Back PU5 > heavy diagonal(16) Diagonal Lag(15) dominates dominates (25) Lead(26) diagonal dominates back (25) diagonal (16) PU1 > heavy

Lag-lag (21|65)

For example, a change from lead-lead to lag-lead would be predicated if:  Phase 5 has a heavier utilization than phase 1  Phases 2 and 5 have heavier total utilization than phases 1 and 6 In this example case, phase 1 would change to a lagging phase, which moves the offset to coincide with the end of phase 2 (most controllers reference the offset to the yellow time of the first coordinated phase) instead of the end of both phases 2 and 6. At this intersection, the offset 58

value does not have to change. However the change to the time that phase 2 will occur during the cycle will change the time that the traffic arrives at the intersection downstream from phase 2 and also the capture efficiency of the traffic that arrives to phase 6 from the intersection upstream of that phase. For example, if phase 2 services northbound traffic and phase 6 services southbound traffic, the intersection to the North will experience traffic arriving earlier in its cycle and the intersection to the South will have the traffic arrive later in its cycle. Offsets are adjusted after the sequence change is evaluated. Similar to the evaluation algorithms for cycle, splits, and offsets, after evaluating the efficiency impact of a potential change to the phase sequence, we next check the effect on the safety by evaluating the regression equation. If the safety is improved, the change is made, otherwise if the safety is degraded, we do not consider the phase sequence change. The evaluation of safety can be enabled or disabled. 4.5 PROTECTED / PERMITTED LEFT-TURN MODE CHANGES (DEFERRED FOR FUTURE CONSIDERATION) The research in Phase 1 of this project showed that the mode of left-turn operations has a significant effect on the safety of an intersection with permitted left turns creating the most conflicts and, not surprisingly, protected left-turns creating the least number of conflicts. Crossing crashes are among the most severe crashes that occur. Efficiency is affected in a slightly different order with permitted being the least efficient, protected being next, and protected-permitted having the highest level of service for the same service volume (assuming the service volume being high enough to require protected-permitted operation). Similarly to the algorithm for selecting phase sequence, the algorithm for modifying left-turn accommodation by use of a table is presented in Table 8. The left-most column lists the current left-turn treatment for a given left-turn. The top row lists the left turn treatment being considered. The cells of the table list the conditions that would justify change from one left turn treatment to the other.

59

Table 8. Rules to evaluate when considering changing left-turn treatment

NEXT LEFT-TURN TREATMENT

CURRENT LT TREATMENT

Permitted only

Permitted only

Protected

N/A

Utilization of left turn lane is heavy

Utilization of left turn phase is very low

N/A

ProtectedPermitted N/A Utilization of leftturn phase is very high

Protected Opposing utilization is low to moderate Utilization of left turn protected portion is moderate to low

N/A Protected-Permitted

N/A

Similarly to the other algorithms, the efficiency measure is evaluated first and then potential detrimental effects on safety are subsequently compared. If the safety is predicted to be degraded, the current operation is retained. Due to practicality constraints of this approach, this algorithm has been postponed for implementation at a later time due to the complexities of recommending changes to this type of operation, and the need for a standard way (or a suite of customized methods) to communicate these changes to controllers, particularly with the emerging standard of flashing yellow arrow versus “doghouse” 5-section heads or other alternative displays used by many agencies (flashing red arrow, etc.). Our discussion with practicing municipal traffic engineers indicated that this type of changes to left-turn operation can be particularly confusing to the public and thus they did not recommend implementation. 4.6 ADAPTIVE CONTROL ALGORITHMS SUMMARY The algorithms could enhance efficiency and safety by accepting changes to the signal timings that only improve both, or are neutral to surrogate safety measures. At the users’ choice, changes to signal timing may not be made if they compromise the surrogate safety measures for the sake of efficiency. The system is not intended to replace or obviate the need for sound traffic engineering but rather supplement the traffic engineer’s toolbox with another tool that can handle fluctuations in demand and short and long-term changes in land use and traffic patterns. Field deployments and studies will be needed to validate that the system improves both safety and

60

efficiency. Innovative approaches may be applied to validate that conflicts are reduced by deploying surveillance cameras and vehicle tracking equipment such as the system developed by UMTRI that has recently been completed (T. Gordon, et. al. 2012). Four principle algorithms were developed for tuning splits, offsets, cycle time, and phase sequence. There are three principle measures of performance used in Kadence: 

Phase utilization. Phase utilization is a surrogate measure of efficiency that represents the degree of saturation of a traffic phase. This measure can be derived directly from the occupancy data measured at stop bar detectors. This measure is used for cycle tuning, cycle selection, split tuning, and phase sequence.



% arrivals on green. % arrivals on green is a measure of efficiency that represents the progression performance of coordinated phases. % arrivals on green can be derived directly from the data measured at upstream detectors on the coordinated phases, and thus requires advanced detectors at 200 feet and preferably 300 feet to work most accurately.



Estimated traffic conflict rate. The total estimated conflicts per hour are a surrogate measure that represents the estimated effect of changing a traffic control parameter on the intersection safety. This measure is a regression model using a feed-forward neuralnetwork that is trained to learn the relationships between signal timing settings and the crossing, lane-changing, and rear-end conflict rates. In the future, the measure may be updated to weight the conflict rates by type to reflect the anticipated higher severity of crossing and lane-change conflicts that result in crashes.

Each of these measures is used in the adaptive control algorithms as detailed in the previous sections. The five optimization stages are executed independently, but in sequence and with the feedback steps as shown in Figure 28. First, (Step 1) the split re-allocation algorithm is executed for each intersection in the system. This identifies if any slack green time can be shifted from one or more phases to another, within the current cycle time, to minimize the maximum phase utilization at the intersection. Surrogate safety measures are evaluated by checking that the reallocation either provides a safety benefit by reducing total conflicts or that the reallocation does not exceed a prescribed threshold. After this re-allocation, the offset adjustment algorithm is executed (Step 2) to identify any modifications to the offsets to improve progression. The total % arrivals on green at both the subject intersection and its neighbors are calculated to represent the efficiency of the proposed change. Similarly to the split calculation, the safety is evaluated by checking that the new offset provides a surrogate safety measures benefit by reducing total conflicts. After the splits and offsets are calculated, modifications to the phase sequence (Step 3) are evaluated with the new split values calculated in Step 1. If any phase sequence modifications are identified that adjust the offset (the departure platoons to adjacent intersections), the offset calculation is re-executed to determine if this change is of further benefit and can be retained, or if the change is detrimental to performance (efficiency or surrogate safety measures).

61

Figure 28. Illustration. Adjustment algorithms flow chart. In the future, it is envisioned that Kadence could evaluate the potential changes to the protected/permitted settings for left-turns (Step 4) using the newly selected phase sequence, if changed in the previous step (at the present time this feature is disabled). If any left-turn settings are deemed to be beneficial for both efficiency and safety, or beneficial for efficiency and within the safety/efficiency trade off value, the split re-allocation algorithm may have to be recalculated, particularly if the left-turn mode is changed from protected to permitted-only. This change in effect omits the left turn phase setting its split to 0 which frees up additional time in 62

the cycle for other phases. It may not be the best policy to simply provide all of that split for its corresponding companion through phase (e.g. phase 6 if the left turn is phase 1). In turn, if the splits are reallocated at this step, the offsets, phase sequence, and protected/permitted settings are re-evaluated as well. Finally, the cycle time adjustment algorithm is evaluated (Step 5). Since cycle time affects all of the intersections in the system, it is important that this adjustment is calculated last, after all of the adjustments/improvements to the individual locations are calculated. As with the phase sequence and protected-permitted settings, if it is deemed beneficial to modify the cycle time, we must re-evaluate the other algorithms within the new value for the cycle (either higher or lower).

63

64

5 SYSTEM PARAMETERS AND SYSTEM USER INTERFACE 5.1 SYSTEM ARCHITECTURE The real-time adaptive efficiency and safety algorithms, known as Kadence, were developed to operate in a centralized/distributed signal system. Its execution was emulated and tested in this project with software Simulation-In-The-Loop (SIL) as illustrated in Figure 29. The ultimate configuration of Kadence in a real-life environment is also illustrated in Figure 30. The algorithms require second-by-second polling of all signal controllers, specifically signal timing and detector inputs. Kadence is developed with an interface to select and configure the system and local signal controller parameters within user-defined signal timing parameter ranges. The system parameters affect Kadence operation and execution of the various tuning algorithms for cycle length, phase utilization, phase sequence selection, offsets, splits, and safety performance. Therefore, users will be required to populate the various parameters based on their ultimate goals to move traffic efficiently and safely.

Figure 29. Illustration. Kadence Software-in-the-Loop Simulation

65

Figure 30. Illustration. Kadence field deployment schematics 5.2 SYSTEM PARAMETERS Kadence has global settings for the system’s parameters which effect all intersections of a network. Figure 31 is a screenshot of the application window for the system parameters, which correspond to cycle length, phase sequence, phase treatment (currently not in use), offset, and split parameters. Users can enable or disable the safety performance algorithm thus limiting its application to efficiency only. System parameters selected under the “Configure Menu” and are defined in the following pages.

66

Figure 31. Illustration. Kadence System Parameters 5.2.1 Cycle Time Parameters Definitions of parameters for cycle length adjustments are listed below. Enable Plan Switching Adjustment: This parameter enables the selection of the next time-ofday (TOD) cycle length earlier and/or later than planned in a TOD schedule. If not enabled, the TOD plan for the cycle selection will be implemented per the user-defined schedule. Cycle Decrease Saturation Threshold: The cycle selection algorithm computes Average Phase Utilization values for each controller. If this average value is found to be less than the Cycle Decrease Saturation Threshold value and the next cycle in the TOD schedule is lower than the current cycle, then the algorithm will switch the system cycle to the next TOD plan. Values for this parameter, namely average phase utilization, range between 0 – 100 percent. Cycle Increase Saturation Threshold: Similar to the Cycle Decrease parameter, if the Average Phase Utilization value is found to be higher than the Cycle Increase Saturation Threshold and the next cycle in the TOD schedule is larger than the current cycle, then the algorithm will switch the system cycle to the next TOD plan. Values range between 0 to 100-percent. Depending on when the Cycle Decrease or Cycle Increase Saturation Thresholds are met, the system cycle may switch earlier or later than planned in the TOD schedule. Min Ints to Jump Forward (minimum number of intersections for cycle to jump forward to next cycle early): This parameter establishes the number of intersections that must have average 67

phase utilization values either higher or lower than the thresholds (depending on if the next cycle in the TOD schedule is higher or lower than the current cycle, respectively) for the next cycle to be initiated earlier than scheduled. It is an integer value. Min Ints to Jump Back (minimum number of intersections to stay in current cycle longer): This parameter establishes the minimum number of intersections that must have average phase utilization values either higher or lower than the threshold (depending on if the next cycle in the schedule is higher or lower than the current cycle, respectively) for the current system cycle to continue to be used past the time when the next cycle in the TOD schedule was originally planned for implementation. It is an integer value. Run Cycle Interval: It is the frequency, in minutes, at which the cycle selection algorithm is evaluated for the entire system. Start Time Cycle Check: It is the number of minutes, before the TOD scheduled plan will change, that the cycle selection algorithm will begin evaluating if a switch to the next cycle is required. End Time Cycle Check: It is the number of minutes, after the TOD scheduled plan would have changed, that the cycle selection algorithm will end evaluating if the TOD plan should stay in the current cycle. Following this period, the next TOD plan will be implemented. Enable Incremental Cycle Adjustment: This parameter enables Kadence to adjust the system cycle length incrementally based on user-defined thresholds. Cycle Increment: It is the interval (in seconds) by which the system cycle length will be incrementally increased or decreased during each cycle evaluation. Min Cycle Time: It is the lowest cycle length (in seconds) to which the system cycle length can be adjusted. Max Cycle Time: It is the highest cycle length (in seconds) to which the system cycle length can be adjusted. 5.2.2 Enable Safety Evaluation Selecting this parameter prompts Kadence to enable the Safety Performance Function (SPF) concurrently with the efficiency optimization. Otherwise, only efficiency is evaluated and tuned. 5.2.3 Phase Sequence Parameters The concept of phase utilization (PhU) is analogous to the degree of saturation concept which also is referred to as the v/c ratio. PhU is calculated using ratios of used green time to available green time. The used green time is calculated from the occupancy data collected from field detectors. Definitions of parameters for phase sequence adjustments are listed below. Heavy Utilization: Expressed in percentage, the phase sequence selection algorithm tests various combinations of phase pairs to determine a sequence change that will optimize

68

efficiency. If a phase utilization value is greater than the Heavy Utilization Threshold, then this phase is determined to have “heavy” or “high” utilization and will be a candidate for further adjustments. Light Utilization: Similar to the heavy utilization parameter, if a phase utilization value (expressed in percentage) is less than the Light Utilization Threshold, then this phase is determined to have “low” or “light” utilization and could prompt the cycle and spilt tuning algorithms for action. Dominate Alpha: This parameter, expressed in percentage, is used to favor which combination of phase utilization values from one pair of phases “dominates” the combination of phase utilization values of other pairs of phases (the pairs that are tested are the compatible through and left-turn phases, such as 1+6, 2+5, 3+8, and 4+7). If (PhU1 + PhU6) > (PhU2 + PhU5 + alpha), then 1+6 “dominates” 2+5, or has “significantly higher” utilization. A value of 40 is suggested when phase utilization is less than 50%. 5.2.4 Protected/Permitted Phase Parameters (currently disabled and subject to future research) Heavy Utilization: Phase utilization (used occupancy / available green time) is the surrogate measure computed for each phase in lieu of the degree of saturation (volume/capacity). The protected-permitted left turn selection algorithm tests various combinations of phase pairs to determine a left-turn type change. If a phase utilization value is greater than the heavy utilization threshold, then this phase is determined to have “heavy” or “high” utilization. Light Utilization (protected – permitted): Similar to the heavy utilization parameter, if a phase utilization value is less than the light utilization threshold, then this phase is determined to have “low” or “light” utilization. 5.2.5 Offset Parameters Max Offset Increment: It is the maximum number of seconds that the offset can be adjusted for each adjustment step. For example, for a maximum value of 5, the offset tuning algorithm is able to evaluate adjustments in the range of [-5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5]. Max Offset Deviation: It is the maximum cumulative adjustment that can be applied to an offset. For example, if the current value of the offset is 30 seconds and the maximum deviation is 10 seconds, then the offset can be changed in the range of [20, 40] but cannot be changed to a value smaller than 20 or larger than 40. Selecting the button “Copy All” causes a transfer of the value to all controllers definition parameters. 5.2.6 Split Parameters Max Split Increment: It is the maximum number of seconds that a split can be adjusted for each adjustment step. For example, if the maximum split increment is set to 4, the split tuning algorithm will select adjustments in the range [-4, -3, -2, -1, 0, 1, 2, 3, 4].

69

Max Split Deviation: It is the maximum cumulative adjustment, expressed in seconds, which can be applied to a split. For example, if the current value of the split is 30 and the maximum deviation is 10, then the split can be changed in the range of [20, 40] but cannot be changed smaller than 20 or larger than 40. The maximum and minimum values for splits are also constrained by the minimum and maximum green times, and the pedestrian crossing times if the “allow oversized peds” parameter is not selected for that phase in the controller menu. Pressing the button “Copy All” causes all system controllers to have the same value. 5.3 CONTROLLER CONFIGURATION Controller parameters are selected and configured for each signal controller that is within the signal network under Kadence’s control. Each signal controller requires configuration individually for detectors, links, phases, ring sequence, timing patterns, and TOD schedule. Screenshots of the signal controller user interface screens are shown below in Figure 32 through Figure 38. Each controller under the Kadence software is defined individually. Figure 32 shows a screenshot of the controller configuration window and the various user-selectable fields.

5 1

2

6 3

7

4

Figure 32. Illustration. Controller configuration parameters The fields in the Controller Definition window are described in the following paragraphs. Field 1: It is an alpha-numeric field where the name (or description) of the intersection is defined. It is typically the name of the cross-street that intersects the major-street.

70

Field 2: The buttons enable and disable the adaptive logic at the controller. The controller will operate in actuated-coordinated mode if disabled. Field 3: The checked boxes indicate the desired adaptive logic algorithms that will adjust the signal timings. Field 4: This text field indicates the Virtual D4 controller file that corresponds to the intersection database. Field 5: This set of parameters comprises the following: 

Controller Number: ID number assigned automatically by Kadence software. The numbering system follows the sequence in which controllers are set up and defined.



Controller Description: Same as the controller name.



Timing: The checked boxes indicate which phases will be favored in the split tuning algorithm.



Biasing: The checked boxes indicate which phases will be prioritized during the split tuning process.



Cycle: The checked boxes indicate which phase utilization is selected to adjust the cycle length tuning algorithm.

Field 6: The checked boxes indicate which combinations of left-turn phase sequence will be considered. There are checkboxes for both mainline (generally, Barrier 1) and the side street (generally, Barrier 2). Field 7: These values are derived from the System Parameters settings but can be altered for individual controllers on a case-by-case basis. 5.4 DETECTORS CONFIGURATION Detector data for an intersection is defined after configuring the controller for the particular location. A screenshot of the detector configuration window is shown in Figure 33.

71

Figure 33. Illustration. Detector configuration parameters The configuration parameters are described below: Detector Num: This is the number assigned to the particular detector within the controller. Description: This field is not being used at this time. Call Phase: This parameter defines the phases that are “called” when the respective detector input is on. This value is automatically populated by the software when the timing values including detector settings are imported from the field controller, using NTCIP 1201 and 1202 or AB3418 standard protocols. Phase Utilization: If the Phase Utilization box is checked then it indicates that it is a stop bar detector and will be used to call phases as well as to calculate phase utilization. Flow Profiling: If the Flow Profiling box is checked then it indicates that it is an advance detector and will be used for phase extension and also for offset tuning.

72

Upstream Controller: The value corresponds to the controller at the upstream intersection. If the detector corresponds to a controller located at the first intersection in the network, then the Upstream Controller is left blank. Downstream Controller: The value corresponds to the controller at the downstream intersection. If the detector corresponds to a controller located at the last intersection in the network then the downstream controller is left blank. Distance Up: This field is the distance, in feet, measured from the trailing edge of the upstream detector to the stop line. Detector Length: Defines the length of the detector. Free Flow Speed: It is the free flow speed on the intersection approach. Time to Flow: This field is the time it takes to travel, at free flow speed, from the upstream detector to the stop line, measured in seconds. Green Time Ext: This field is not being used at this time. Sec to Brake Point: This value is automatically computed and populated by the software when the timing values including detector settings are imported from the field controllers. The “Sec to Brake Point” is an adjustment to the “Time to Flow” and it is defined as the time measured from when a driver starts to break after passing the upstream detectors to the time the driver can stop safely at the stop line. This parameter can effect on offset turning. 5.5 LINKS CONFIGURATION This section describes the link configuration parameters. In Kadence, the links connect two adjacent signalized intersections that should be coordinated. A screenshot of the user interface is shown in Figure 34. In the figure, Field 1 is a table in which the rows represent the “from” intersection and the columns represent the “to” intersection. Once a combination of “from” and “to” is selected – by clicking “UND” –, Field 2 shows the scheme of the link. The directional link will show the name of the upstream and downstream intersections. Field 3 is populated with data that corresponds to the phase number of the selected link direction, the distance between the two selected intersections, and the free flow speed on the link.

73

1

2 3

Figure 34. Illustration. Link definition parameters 5.6 PHASING DATA The existing phasing information for each controller is imported directly from the field controllers using NTCIP 1201 and 1202 or AB3418 standard protocols, and cannot be modified manually through the software. Figure 35 shows the typical phasing data representation in Kadence for an individual controller.

74

Figure 35. Illustration. Phase configuration parameters

5.7 RING SEQUENCES The existing ring sequence information for each controller is imported directly from the controller database and cannot be modified through the user interface. Figure 36 shows the ring sequence for a given controller.

Figure 36. Illustration. Ring-barrier sequence The table in the screenshot describes the information for each ring. The letters “a” and “b” under Ring Sequence represent the first and second barrier, respectively.

75

5.8 TIME-OF-DAY (TOD) PATTERNS The data regarding the TOD patterns is also imported directly from the field for each controller, using NTCIP 1201 and 1202 or AB3418 standard protocols. Figure 37 depicts how the data is shown in Kadence.

Figure 37. Illustration. Pattern definition parameters 5.9 PATTERNS SCHEDULE The patterns schedule is also imported directly from the controller database. Figure 38 represents a screenshot of the data as it appears in Kadence. It highlights when each signal timing pattern begins and ends.

76

Figure 38. Illustration. TOD plan parameters 5.10 VISSIM SIMULATION PARAMETERS (SIMULATION-IN-THE-LOOP) The performance of the Kadence’s real-time adaptive efficiency and safety algorithms was tested using Software-In-The-Loop (SIL) simulation. SIL simulation is a methodology typically used to test signalized operations at intersections in a simulated environment using a virtual replica of a real-world traffic controller firmware. A typical SIL setup comprises of a microscopic traffic simulation tool and a virtual traffic controller software which allows the simulation tool to communicate with the virtual controller and program various signal timing parameters. The research team elected to utilize the microscopic simulation software VISSIMTM in conjunction with the 4th dimension virtual D4 controller software for the SIL. VISSIMTM (Verkehr In Städten – SIMulationsmodell), which translates to “Traffic in cities simulation model”, is a time-step and behavior based microscopic traffic simulation software (PTV AG. 2008). It is characterized as a microscopic simulation model because of its ability to analyze each entity of the network at an individual level. VISSIM is capable of simulating

77

multiple modes of traffic including (but not limited to) cars, HGVs, HOVs, bus transit, light rail, heavy rail, rapid transit, cyclists and pedestrians, for rural as well as urban conditions. The VISSIM model consists internally of two distinct components, the traffic simulator and the signal state generator, which constantly communicate detector calls and signal status with each other through an interface. The traffic simulator is a microscopic traffic flow simulation model including car following and lane change logic. The signal state generator is signal control software polling detector information from the traffic simulator on a discrete time-step basis (down to 1/10 of a second). It then determines the signal status for the following time-step and returns this information to the traffic simulator. This interaction is the driving force behind modeling a signalized intersection. VISSIM version 5.12 (release 14) was used in this research for simulating the test cases. The Virtual D4 Suite is a comprehensive set of software tools for SIL traffic signal controller simulation for use with VISSIM. With these tools, Signal Control Junctions (SCJs) within a VISSIM model can be configured to run Virtual D4 traffic signal controllers using NTCIP or other standard communications protocol for traffic controllers such as AB3418e. Figure 39 shows a flowchart of the SIL setup. TRAFFIC SIMULATION SOFTWARE (VISSIM)

VIRTUAL D4 CONTROLLER INTERFACE

VIRTUAL D4 CONTROLLER DATABASE EDITOR

Figure 39. Illustration. SIL Data Flow 5.10.1 Model Geometry The baseline test case scenario modeled in VISSIM consisted of a three-intersection corridor. The intersections were named as intersection M (middle), intersection W (west), and intersection E (east). The following highlights the basic assumptions made with respect to the geometry: 

All the lanes within the study network were assumed to be 12 feet wide.



The Major Street (Main Street) is oriented in the east-west direction.



Minor streets (Side Street) are oriented in the north-south direction.



Arterial spacing between the intersections W and M, and M and E is 1200 feet and 1000 feet, respectively.



Three through lanes on the major and side street approaches at all three intersections.



Dedicated left-turn and right-turn lanes (250 feet long) were assumed for the major street and side street approaches at all three intersections.

78



Free-flow speeds along the major street and side street approaches are 50 mph and 35 mph, respectively.

Figure 40 shows a snapshot of the VISSIM network.

Figure 40. Illustration. VISSIM Study Network 5.10.2 Traffic Demand Data The following describes pertinent assumptions made with respect to generating traffic demand data: 

The base line case scenario assumed an average volume-to-capacity ratio between 0.95 and 1.00 to reflect a high demand scenario, and between 0.50 and 0.60 to reflect a low demand scenario for all movements.



Traffic composition consists of 98% passenger cars and 2% heavy vehicles.



A uniform Peak Hour Factor of 0.92 was adopted for all movements.



No pedestrian, bicycle or transit volume is considered for the purpose of this research.

5.10.3 Traffic Control Data The following assumptions were made with respect to traffic control at the three signalized intersections: 

All intersections operate in a fully actuated-coordinated mode. 79



All intersections operate at the same cycle length. Two TOD patterns were assumed for testing purposes: a cycle of 70 seconds (for low traffic demand) and a cycle of 130 seconds (for high traffic demand).



Signal phases were assigned for all movements using NEMA phasing convention, wherein phases 2 and 6 denote the major street through movements and phases 4 and 8 denote the minor street through movements.



Left turn treatment was set to protected-only at all intersections. Left turn sequence was set to lead or lag and varied for each approach.



The change and clearance intervals (Yellow + All Red times) for all phases were calculated as per the methodology outlined by the Institution of Transportation Engineers (ITE). However, the All Red (AR) time was limited to 1 second in the case of left turns, as the calculated value of 2.5 seconds was deemed to excessive.



A minimum green time of 5 seconds for left turns and 10 seconds for the through movement was assumed on the Main Street. A minimum green time of 7 seconds was assumed for the side street phases.



Advanced detection was placed along the major street at all intersections assuming a 5 seconds passage time. The detectors placement was set to 330 feet from the stop line.

SYNCHROTM traffic software was utilized to generate the base line optimum signal timing parameter including cycle, split, offset and phase sequence. The optimized timing data was then imported into the Virtual D4 controller to run in conjunction with the VISSIM traffic simulation. The Virtual D4 Suite User’s Manual has instructions on setting up and fine tuning the Virtual D4 controller parameters. 5.10.4 Simulation Parameters For the duration of the SIL testing, the simulation parameters in VISSIM were kept unchanged and consistent with the following: 

Simulation period: 4500 seconds (for single TOD cycle scenario) and 8100 seconds (for two-TOD cycle scenario). This comprised of a mandatory 900-second “seeding” period wherein no performance measures were recorded.



The simulation resolution was set to 10 time step(s)/simulation second. This matches with the resolution of the Virtual D4 controller.



Simulation speed was set to 2.0 simulation seconds/real world second.

5.11 SYSTEM PERFORMANCE OUTPUT Throughout the course of a SIL simulation, Kadence generates an extensive output file in which it keeps a log of the various statistics used during the process. This log file is updated in realtime during the simulation and displays adaptive statistics for split, cycle, offset, phase sequence, 80

efficiency and safety (if enabled) data. The top of the log file lists all Kadence’s configuration data: system parameter values, algorithm settings, controller settings and detector states, TOD pattern schedule, and cycle/split/offset information for separate plan events. This enables users to check and verify the settings before proceeding with a simulation run. The log file is updated on a minute-by-minute basis with statistics used during the Kadence’s decision making process. The log file can be opened and viewed using any text editor tool. Figure 41 shows a fragment of the log file showing the start of the adaptive logic process and detector information of a given controller.

1 3

2 8 4

5

6

7

8

Figure 41. Illustration. Typical Kadence log file illustration The highlighted information in Figure 41 is explained as follows: 

Field 1: PC TIME is the date and time the simulation was performed. SIM-TIME is the instant during the simulation that adaptive logic process started.



Field 2: It indicates the controller being updated next.



Field 3: It lists the detectors that are connected to the controller indicated in Field 2. The detectors are listed in the order that they are first called in the simulation.

81



Field 4: It indicates the time-of-day pattern that is in use.



Field 5: It is the phase number in Ring 1.



Field 6: It indicates the interval of the phase in Ring 1 (values are “1” for green, “6” for yellow, and “10” for red).



Field 7: It is the phase number in Ring 2.



Field 8: It indicates the interval of the phase in Ring 2 (values are “1” for green, “6” for yellow, and “10” for red).

Once Kadence sends the new data to the controller, it also plots the information on graphs displayed in a separate window. Plots of adjusted timing and phasing data are displayed continuously during the simulation period. Figure 42 illustrates the output graphs generated in Kadence.

82

1 2 1

6 1

7 8

3

9 4

5

Figure 42. Illustration. Typical output graphs The individual options and graphs depicted in under the highlighted fields are described as follows: 

Field 1: The “Select Log File” button allows users to select a specific log file pertaining to a single instance or operational scenario. A copy of the log file can also be saved using the “Save As” button.



Field 2: Individual controllers within the system can be accessed from the dropdown menu. Once selected, the performance data for a particular controller can be viewed by pressing the “Generate” button. The performance data may also be exported to a .CSV file, by pressing the “Export CSV” button, to be viewed in Microsoft Excel.

83



Field 3: Displays the adjustment made to the offset setting in the selected controller during the course of the simulation time.



Field 4: Displays the adjustments made to cycle length (both “Plan Switching” as well as “Incremental” type) at the particular location.



Field 5: Displays adjustments made to the phase sequences at the selected location.



Field 6: Users can select whether the interface displays the parameter adjustment graphs or the safety performance graphs. This information is displayed only when the user enables the “Safety Performance Function”.



Field 7: This option allows users to select the X-Axis parameter of the output graphs. The available options are “Simulation Time” or “Command Interval”.



Field 8: Depicts the adjustments made by Kadence to splits in Ring 1 only.



Field 9: Depicts the adjustments made by Kadence to splits in Ring 2 only.

5.11.1 Offset Adjustment Outputs During the offset tuning process, Kadence outputs the candidate offset values and its estimated total captured flow. The candidate offset with higher captured flow is sent to the controller. However, if SPF is enabled, Kadence only sends the new data to the controller if the predicted total conflicts (with the proposed offset) are lower than the existing total conflicts. Figure 43 is a sampler of the offset adjustment’s log from Kadence.

2 1

3 4

Figure 43. Illustration. Offset adjustment log (with SPF enabled)

84

The output information depicted in Figure 43 under the highlighted fields is explained as follows: 

Field 1: These values are the offsets considered during the adjustment process.



Field 2: It represents the MOE calculated for each candidate offset. The MOE is the sum of captured flow from all approaches.



Field 3: It indicates the total number of conflicts estimated for the actual offset and for the offset with higher total captured flow.



Field 4: It indicates the offset selected by the algorithm.

The graphical output of the offset tuning is illustrated in Figure 44. The topmost image shows the offset changes along the simulation run. The bottom one shows the number of predicted and current estimation of conflicts at the instant of the given adjustment.

85

OFFSET = 57 s

OFFSET = 48 s

Sim-Time 8:53:44

Number of Conflicts estimated for actual offset (48 s) = 204

Number of Conflicts estimated for selected offset (57 s) = 195

Sim-Time 8:53:44

Figure 44. Illustration. Offset adjustment and safety prediction output graphs

86

5.11.2 Phase Sequence Adjustment outputs Kadence adjusts the phase sequence taking into consideration the calculated utilization of phase pairs. If SPF is enabled, Kadence elects to make adjustments only when the predicted total conflicts (with the proposed phase sequence) are lower than the existing total conflicts. The output information depicted in Figure 45 illustrates the decision process of the phase sequence adjustment. The process is performed for each barrier (phase group) separately.

1

2 3

4

5

6

Figure 45. Illustration. Sequence adjustment log The information under the highlighted fields is explained as follows: 

Field 1: The number in the parenthesis is the phase number for the referred phase group.



Field 2: This number is the phase utilization, in percentage, that corresponds to the phase located to its left side.



Field 3: It indicates the left-turn phase sequence (lead-lead, lead-lag, lag-lead, lag-lag).



Field 4: It indicates the result of the sequence switching logic: “both rings swap” which means that both left-turn phases are switching, “Ring 1 swap”, “Ring 2 swap”. Also, it may return a “No sequence change” if the phase utilization does not reach the thresholds.



Field 5: It indicates the phase utilization percent and how it compares to the thresholds.



Field 6: It indicates which approach or left-turn phase have the higher phase utilization.

The graphical representation of the phase sequence adjustment as it is presented by Kadence’s user interface is depicted in . The vertical axis represents the sequence that phases are called. Sequences numbered 1 through 4 correspond to ring 1; sequences 5 through 8 refer to ring 2.

87

Phases switching positions in the ring sequence.

Ring 2

Ring 1

Figure 46. Illustration. Sequence adjustment graph 5.11.3 Split Adjustment Outputs Kadence adjusts the splits depending on the utilization values calculated for each phase. In addition, if SPF is enabled, Kadence elects to make adjustments only when the predicted total conflicts (with the proposed split) are lower than the existing total conflicts. Figure 47 illustrates the split adjustment indicated in the log file.

1

2 1

3

4

5

7

6

8

. Figure 47. Illustration. Split adjustment log The information under the highlighted fields is explained as follows:

88



Field 1: The strings of characters “|a|” and “|b|” indicates the ring-barrier divider.



Field 2: It is the phase number.



Field 3: It is the actual split (current at that instant of the simulation).



Field 4: This value represents the phase utilization.



Field 5: It indicates the action of the algorithm based on the phases’ utilization. In this particular example, it is taking time from phase 1 is added to phase 2. The values in the parenthesis are the before and after phase utilization of both phases.



Field 6: These are the minimum and maximum bounds for the given phase split and its actual value at that instant. The values are determined based on the maximum split increment and deviation that were defined in the System Parameters menu. The split candidate is also limited by the minimum split established for each movement.

Figure 48. and Figure 49 show the graphical representation of the trade-offs among phases for ring 1 and ring 2, respectively, during the split adjustment process. Figure 50 depicts the instant during the simulation when the splits are adjusted and, since the safety function is enabled, it shows the reduction in the predicted number of conflicts.

Figure 48. Illustration. Split adjustment Ring 1

89

Figure 49. Illustration. Split adjustment Ring 2

89 conflicts are estimated before the adjustment.

87 conflicts are predicted to occur after the adjustment.

Figure 50. Illustration. Split surrogate safety measures prediction graph

90

5.11.4 Cycle Selection Outputs Kadence adjusts the cycle length based on the average phase utilization calculated for each controller. However, unlike the splits or offsets, the cycle length is merely switched to the next higher (or lower) cycle in the TOD schedule depending on saturated (or unsaturated) thresholds being met. Cycle length may also be adjusted on an incremental basis if this option is selected at the System Parameters settings. Figure 51 illustrates the cycle adjustment indicated in the log file. The cycle length may be incremented based on a user-specified value such as 1 second to 10 seconds.

1

2

3

Figure 51. Illustration. Log illustration of plan switching from higher to lower cycle The highlighted fields as described as follows: 

Field 1: This indicates the phase utilization value.



Field 2: This is actual planned cycle length running in the controller.



Field 3: This is the next plan and cycle length selected by Kadence that will improve performance and safety (if enabled).

Figure 52 shows the graphical representation of the cycle length adjustment using the TOD plan switching algorithm. In this case, the actual cycle length running is 150 seconds and Kadence switches it to the next plan that has a cycle of 130 seconds. Figure 53 shows that the cycle length is adjusted by increments of time that are added or withdrawn from the actual cycle to optimize the performance and safety (if enabled).

91

Plan 4 150 seconds

Plan 5 130 seconds

Figure 52 . Illustration. Graphical illustration of plan switching from higher to lower cycle

5s increment 5s increment

Figure 53. Illustration. Graphical illustration of incremental adjustment from higher to lower cycle

92

5.12 SUMMARY Kadence is a real-time adaptive efficiency and surrogate safety measures algorithm platform, developed to operate in a centralized/distributed signal system. Kadence is designed with a userfriendly interface to configure the system and local signal control parameters. Kadence provides detailed log files about the decision-making process and displays graphical outputs of adjusted signal timing parameters. Kadence allows the user to enable or disable safety evaluation on either a system wide or a local level, i.e. by intersection. Thus, the user can exclude certain intersections from the safety evaluation if so desired. Kadence operates only as an efficiency tuning model if the safety performance function is not enabled in the configuration menu. The efficiency mode enables tuning of the cycle length, cycle selection, splits and offset, and left-turn phase sequence. When operating Kadence with the SPF enabled, signal timing changes can only be implemented in the system if the SPF resulted in equal or a lesser number of conflicts. When Kadence is operating with the SPF disabled, than changes in signal timing derived from the efficiency algorithms will always be implemented regardless of their effect on safety.

93

94

6 ALGORITHMS VERIFICATION 6.1 OVERVIEW This chapter describes the process followed to analyze and verify the performance of the Kadence adaptive logic algorithms. The process entailed 11 different case studies derived from a single baseline model of three signalized intersections. The process involved the use of three analysis tools (Synchro, VISSIM and SSAM), virtual D4 traffic signal controller software, and a VISSIM Vehicle Actuated Programming (VAP) module produced by Fourth Dimension Traffic. The virtual D4 controller software emulates an NTCIP-compliant signal controller, which could be used with Kadence applications in the field or in a laboratory setting. The process of developing the verification cases, and the interpretation of the results are discussed in the following section. 6.2 SIMULATION METHODOLOGY The objective of the verification process is to create various test cases with parameters that will properly invoke the various parameter-tuning logics (cycle, split, offset and phase sequence) in real-time, in Kadence, and observe the end results to determine if they are logical. Therefore, test cases were created to test each algorithm individually. Also, testing was also performed with various combinations of the different parameter-tuning algorithms. The cases were structured so that the network is stressed near saturation, specifically with intersections operating at levels of service “E” or worse, but none of the intersections V/C ratios exceeded 1.20. A base line model was developed and used as a common base line for all test cases. Three additional analysis scenarios were performed and compared to the base line model. All scenarios are discussed below as follows: 6.3 TEST SCENARIOS 6.3.1 Baseline Arterial Model The baseline scenario consists of a corridor with three signalized intersections on a major street. The intersections are named as intersection M (middle), intersection W (west), and intersection E (east). Figure 54 shows a schematic of the baseline model.

95

Figure 54. Illustration. Baseline arterial configuration Geometry: The main street is a six-lane divided arterial and runs east-west. The side streets have three approach lanes and run north-south. All approaches at all intersections have single exclusive left and right-turn lanes. Speed limits on the main and side streets are 50 mph and 35 mph, respectively. Detection: Stop bar detectors are placed on all lanes at the intersections as shown in Figure 55. Advanced detection is placed on the approach of all three lanes. Stop bar detectors are 32 feet long and advanced detectors are 6 feet long, and placed 315 feet upstream from the stop bar.

Stop Bar Detectors

Advanced Detectors

Figure 55. Lane configuration and detection location.

96

Traffic Volumes: Volumes were on all intersections were selected to maintain a volume-tocapacity ratio (v/c) close to 1.0, but not greater than 1.20. Signal Timing and Phasing: All intersections are fully actuated and operate in NEMA 8-phase operation. The phase sequence at intersections W and E consists of a lead-lag left-turn, and lag left-turns at intersection M. All left-turns are protected. The baseline model runs a single TOD pattern, optimized for the best cycle length and splits. The minimum green is 10 seconds for the main streets, and 7 seconds for the side streets. The change and clearance intervals, shown in Table 9 are calculated in accordance with prescribed method of the ITE Manual of Signal Timing. Table 9. Change and clearance intervals for the baseline model. Clearance Interval

Main Street

Side streets

Yellow

4.5 seconds

3.5 seconds

All Red

1.0 seconds

2.0 seconds

Figure 56 shows the offsets, approach volumes, cycle length, and phase sequence for all intersections.

Figure 56. Illustration. Demand and signal timing for the baseline scenario

97

Table 10 and Table 11 list Kadence’s System Parameters and Controller settings, respectively. Table 10. Kadence system parameters used for verification process. Parameter

Value

Cycle Decrease Saturation Threshold Cycle Increase Saturation Threshold Min Intersections to Jump Forward Min Intersections to Jump Back Run Cycle Interval Start Time Cycle Check End Time Cycle Check Cycle Increment Min Cycle Time Max Cycle Time Heavy Utilization Light Utilization Dominate Alpha Max Offset Increment Max Offset Deviation Max Split Increment Max Split Deviation

50 85 2 2 10 minutes 60 minutes 60 minutes 10 seconds 70 seconds 130 seconds 85 % 40 % 5% 10 seconds 45 seconds 10 seconds 45 seconds

Table 11. Signal Controller parameters used for verification process Parameter Adjust Splits - Timing

Value

Deviation – Max Offset

All phases are set to be adjusted Coordinate phases have biased splits (Phases 2 and 6) Only mainline phases are accounted for cycle adjustments (Phases 1,2, 5 and 6) Only mainline phases are involved on Phase Sequence adjustments (Phases 1, 2, 5 and 6) 45 seconds

Deviation – Max Split

45 seconds

Adjust Splits - Biasing Adjust Splits – Cycle Phase Sequence

Outputs from the baseline model as shown in tabulated summaries and graphical illustrations are identified under the Synchro-VISSIM scenario and compared to the other three scenarios.

98

6.3.2 Analysis Scenario 1 (VISSIM-Virtual D4) This scenario is a derivation of the baseline model. It is structured so that changes in actuated and coordinated signal parameters and traffic demand are imposed on the existing optimized solution (baseline), without re-optimizing the network. This scenario is simulated as an actuated coordinated network using VISSIM and Virtual D4 controller software. 6.3.3 Analysis Scenario 2 with VISSIM-Virtual D4-Kadence (SPF disabled) This scenario is a derivation of scenario 1 and is tailored to address only efficiency since the safety performance function under this scenario is disabled. The VISSIM and Virtual D4 configuration parameters in this scenario are identical to scenario 1.The adaptive logic and tuning algorithms of Kadence are used to modify the existing signal timing parameters in realtime. 6.3.4 Analysis Scenario 3 with VISSIM-Virtual D4-Kadence (SPF enabled) This scenario is a derivation of scenario 1 and is tailored to address both efficiency and safety. The VISSIM and Virtual D4 configuration parameters in this scenario are identical to scenario 2. In addition to the safety performance function evaluation, the adaptive logic and tuning algorithms of Kadence are used to modify the existing signal timing parameters in real-time. Each analysis scenario included one seeding interval of 900 seconds, which was excluded from the analysis,, and an average of five runs of each scenario. Case studies created to test offsets, splits and phase sequence algorithms included four analysis intervals for a total of 3600 seconds. Case studies created to assess the cycle length algorithm included eight analysis intervals for a total of 7200 seconds. Performance output data such as delays, travel time, stops, throughput and total number of conflicts were acquired and summarized. VISSIM trajectory data was processed in SSAM, and the total number of conflicts was generated from Kadence using the SPF index 6.4 TESTING CASES AND RESULTS The algorithm verification process involved the creation, simulation and output processing of 11 test cases based on variable traffic demand conditions and signal timing parameters. In the first 11 test cases, as shown in Table 12, each case was configured to test one of the tuning algorithms, and starting with a non-optimum solution. However, a combination of algorithms was also performed for cases 4 and 11. Simulations of each case were performed for the three analysis scenarios listed earlier and also compared to the baseline optimized results.

99

Table 12. Test cases and respective adaptive algorithm Case No.

Case Description

Tuning Algorithm

1

Non-Optimum Offset Selection At Intersection M

Offset

2

Non-Optimum Offset Selection At All Intersections

Offset

3

Over-Saturated Left-Turn (EB) At Intersection M

Split

4

Over-Saturated Left-Turns (EB And WB) At Intersection M

Split

5

Non-Optimum Phase Sequence Selection At Intersection M - EB Lead

Sequence

6

Non-Optimum Phase Sequence Selection At Intersection M - WB Lead

Sequence

7

Non-Optimum Phase Sequence Selection At Intersection M - Both Lead

Sequence

8

Higher Demand Starts Late

Cycle

9

Higher Demand Starts Early

Cycle

10

Lower Demand Starts Late

Cycle

11

Lower Demand Starts Early

Cycle

These cases were structured to test each algorithm individually. However, some conditions were also simulated while enabling more than one adjustment algorithm at a time. During the VISSIM simulations, output data was collected and output in text reports. The output data is comprised of node evaluation, network performance, travel times, and vehicle trajectory. All data collection starts after the simulation seeding period (900 seconds) is completed. Node Evaluation files (*.kna) present intersection-specific data collected using user-defined node boundaries within the VISSIM network. Node evaluation data is collected and summarized for every 900-second interval. Nodes are defined as shown in Figure 57. Data collected in nodes are listed in Table 13.

100

NODE 1

NODE 2

NODE 3

Figure 57. Illustration. Node locations.

Table 13. Node evaluation data. Performance Measures Delay Queues Stops

Description

Unit

Average Delay per Vehicle Average Queue Length Average Number of Stops per Vehicle

Seconds Feet -

Network performance output files (*.npe) consist of several parameters that are aggregated for the total simulation period and the entire network. Network performance data collected during simulation is listed in Table 14. Table 14. Network Performance data. Performance Measures Average Delay Time per Vehicle Average Number of Stops per Vehicle Average Speed Total Delay Time Number of Stops Total Travel Time

101

Unit Seconds Miles per Hour Hours Hours

Travel time output files (*.rsz) encompass data for each travel time section defined in the network summarized for each 900-second interval. Travel times segments were defined along the mainline and are depicted in Figure 58.

WESTBOUND

EASTBOUND

Figure 58. Illustration. Travel time segments. Vehicle trajectories were collected along all link segments during all simulation periods. The trajectory files were processed in SSAM to determine the number of total conflicts. Postprocessing filtering was established. Table 15 lists the filtering parameters in SSAM. Minimum TTC and PET were set to 0.1 seconds rather than zero in order to remove all crashes resulted from the simulation runs. Only conflicts along the main street approaches and intersections were accounted for since all changes imposed in the analysis was focused on main street movements. Table 15. Filtering thresholds for conflicts in SSAM. Parameter

Value

Maximum time-to-collision (TTC)

1.5 seconds

Maximum post-encroachment time (PET)

5.0 seconds

tMinTTC (Simulation Time)

scenarios without cycle adjustment algorithm scenarios with cycle adjustment algorithm

900 s ≤ tMinTTC ≤ 4500 s 900 s ≤ tMinTTC ≤ 8100 s

Rear-end Angle

30 degrees

Crossing Angle

80 degrees

102

The case description and the results of the simulations are discussed below. 6.4.1 Case 1, Non-optimum offset selection at intersection M This case presents a non-optimum offset selection at intersection M. The corridor layout, traffic volumes, and all remaining signal timing and phasing parameters are the same as the baseline case. Figure 59 shows schematics of the traffic characteristics for the above referred scenario and depicts in a shadowed area the parameter that is modified. In this case, the offset at intersection M is set to 88 seconds. The baseline optimum offset is 63 seconds.

Figure 59. Illustration. Demand and signal timing for Case 1. The objective of this test case is to verify the functionality of the offset tuning algorithm for a single intersection in the arterial, by setting a non-optimum offset to intersection M and constraining all other signal timing settings at the remaining two intersections. The results from the simulations, as shown in Table 16, indicate that Kadence is able to find a favorable solution that would meet both the efficiency and safety objectives of this research. It is expected that Kadence would find a better solution than is shown in Table 16 if the signal timing parameters at the adjacent intersections were not constrained and allowed to be tuned.

103

Table 16. Case 1 results - MOE collected at intersection M (through lanes only) SynchroVISSIM

VISSIMVirtual D4

Kadence Safety Disabled

Kadence Safety Enabled

900 - 1800

46.5

36.8

33.3

36.8

1800 - 2700

45.1

37.8

35.4

37.8

2700 -3600

43.8

35.5

32.1

35.5

3600 – 4500

44.3

38.5

29.7

38.5

900 - 1800

27.4

25.5

19.1

25.5

1800 - 2700

30.8

29.0

21.0

29.0

2700 -3600

31.1

30.5

19.0

30.5

3600 – 4500

48.8

28.8

25.2

28.8

17

15

16

15

7620 / 7498

7620 / 7594

7620 / 7565

7620 / 7593

Average Delay Westbound (seconds)

Simulation Interval

Average Delay Eastbound (seconds)

Simulation Interval

MOE

Total Conflicts (per 1,000 vehicles) Vehicles (input/simulated)

6.4.2 Case 2, Non-optimum offset selection at all intersections This case presents non-optimum offset selection at all three intersections. The corridor layout, traffic volumes, and all signal timing and phasing parameters remain the same as the baseline case. Figure 60 shows a schematic of the traffic characteristics for the above referred scenario and highlights the parameters that are modified. In this case, the offsets at intersections W, M and E are set to 53 seconds, 88 seconds, and 101 seconds, respectively. The baseline optimum offsets are 78 seconds, 63 seconds, and 76 seconds, respectively. The objective of this test case is to verify the functionality of the offset tuning algorithm for multiple intersections, and determine if Kadence is capable of providing a more efficient and safer solution than the baseline model’s solution. The results in Table 17 show improvement in mainline travel time, only in the eastbound direction. Travel time results for the westbound direction are marginal. However, significant improvement (over 40-percent) in the vehicle throughput and the total number of conflicts was noted in the Kadence results.

104

Figure 60. Illustration. Demand and signal timings for Case 2

Table 17. Case 2 results - MOE collected mainline arterial SynchroVISSIM

VISSIMVirtual D4

Kadence Safety Disabled

Kadence Safety Enabled

900 - 1800

180.2

200.7

187.0

197.7

1800 - 2700

193.5

219.8

202.3

195.0

2700 -3600

183.7

236.6

176.1

174.5

3600 – 4500

189.8

224.6

203.8

178.4

900 - 1800

165.7

197.8

196.9

197.4

1800 - 2700

173.7

238.7

195.2

201.9

2700 -3600

168.3

288.5

212.3

199.1

3600 – 4500

208.1

315.5

243.6

224.4

117

135

77

72

3554/2440

3554/2333

3554/3859

3554/3888

Travel Time Westbound (seconds)

Simulation Interval

Travel Time Eastbound (seconds)

Simulation Interval

MOE

Total Conflicts, EB Left-Turn (per 1,000 vehicles) Vehicles (input/simulated)

105

6.4.3 Case 3, Over-saturated left-turn (EB) at intersection M This case presents an increase in the demand of left-turn vehicles at intersection M along the eastbound direction. The corridor layout and traffic volumes at the other two intersections and remaining approaches of intersection M are kept the same as the baseline model. Also, all signal timing and phasing parameters remain the same as the baseline case. Figure 61 shows a schematic of the traffic characteristics for this test scenario, and highlights the eastbound left turn volume at intersection M, which is increased from 200 vph to 340 vph.

Figure 61. Illustration. Demand and signal timings for Case 3. The objective of this test case is to verify the functionality of the split adjustment algorithm, and determine if Kadence is capable of altering the phase splits in real-time to accommodate the increase in traffic volume for the eastbound left-turn volume at intersection M. As shown in Figure 62 Kadence adjusted the phase splits favorably and reallocated time from phase 6 (eastbound through) to phase 5 (eastbound left). The results in Table 18 also demonstrate that Kadence reduced delays significantly.

106

Phase 6

Phase 5

Figure 62. Split adjustments for Case 3. Table 18. Case 3 results – MOE collected at intersection M SynchroVISSIM

VISSIMVirtual D4

Kadence Safety Disabled

Kadence Safety Enabled

900 - 1800

63.4 (E)

125.0 (F)

81.2 (F)

81.2 (F)

1800 - 2700

67.5 (E)

223.6 (F)

84.5 (F)

87.9 (F)

2700 -3600

86.0 (F)

371.5 (F)

107.3 (F)

111.9 (F)

3600 – 4500

128.4 (F)

646.2 (F)

166.5 (F)

151.4 (F)

900 - 1800

77.1 (E)

143.0 (F)

83.7 (F)

83.7 (F)

1800 - 2700

64.5 (E)

195.9 (F)

55.1 (E)

55.3 (E)

2700 -3600

74.9 (E)

231.1 (F)

73.3 (E)

76.7 (E)

3600 – 4500

59.2 (E)

253.8 (F)

195.9 (F)

182.6 (F)

191.6

143.6

133.0

127.9

205/206

310/249

310/266

310/269

Average Delay WBL (seconds)

Simulation Interval

Average Delay EBL (seconds)

Simulation Interval

MOE

Total Conflicts, EB Left-Turn (per 1,000 vehicles) Vehicles, EB Left-Turn (input/simulated)

107

6.4.4 Case 4, Over-saturated left-turns (EB and WB) at intersection M This case presents an increase in the demand of left-turn vehicles at intersection M along the eastbound and westbound directions. The corridor layout, initial signal timing and traffic volumes at the other intersections and remaining approaches of intersection M are kept the same as the baseline case. Also, all remaining signal timing and phasing parameters remain the same as the baseline case. Figure 63 shows a schematic of the traffic characteristics for this test case and highlights the east and westbound left turn volumes at intersection M. The left-turn demand at intersection M is increased from 200 vph to 340 vph for eastbound and from 185 vph to 310 vph for the westbound, respectively.

Figure 63. Illustration. Demand and signal timings for Case 4 The objective of this test case is to verify the functionality of the split adjustment algorithm and determine if Kadence is capable of altering the phase splits in real-time to accommodate the increase in traffic volume for the eastbound and westbound left-turn volumes at intersection M. As shown in Figure 64, Kadence adjusted the phase splits favorably and reallocated time from phases 2 and 6 to phases 1 and 5. The adjustments, as shown in Table 19, also resulted in reduction to delays, total number of conflicts and increased throughput for both movements.

108

Figure 64. Illustration. Split adjustments for Case 4

109

Table 19. Case 4 results - MOE collected at intersection M (EBL and WBL) SynchroVISSIM

VISSIMVirtual D4

Kadence Safety Disabled

Kadence Safety Enabled

900 - 1800

63.4 (E)

125.0 (F)

81.2 (F)

81.2 (F)

1800 - 2700

67.5 (E)

223.6 (F)

84.5 (F)

87.9 (F)

2700 -3600

86.0 (F)

371.5 (F)

107.3 (F)

111.9 (F)

3600 – 4500

128.4 (F)

646.2 (F)

166.5 (F)

151.4 (F)

192

144

133

128

225/209

340/230

340/292

340/291

900 - 1800

77.1 (E)

143.0 (F)

83.7 (F)

83.7 (F)

1800 - 2700

64.5 (E)

195.9 (F)

55.1 (E)

55.3 (E)

2700 -3600

74.9 (E)

231.1 (F)

73.3 (E)

76.7 (E)

3600 – 4500

59.2 (E)

253.8 (F)

195.9 (F)

182.6 (F)

150

130

126

119

205/206

310/249

310/266

310/269

Average Delay EBL (seconds)

Simulation Interval

MOE

Total Conflicts EB Left-Turn (per 1,000 vehicles)

Average Delay WBL (seconds)

Simulation Interval

Vehicles EB Left-Turn (input/simulated)

Total Conflicts, WB Left-Turn (per 1,000 vehicles) Vehicles, WB Left-Turn (input/simulated)

This case was also tested with a combination of offsets, phase sequence and split adjustments enabled concurrently. Table 20 shows the Kadence efficiency-based results, while Table 21 shows both efficiency and safety-based results with the SPF being enabled. The findings clearly demonstrate that Kadence’s best performance is when all parameters are tuned concurrently. Table 20. EB Left-turns - Average Delays per Vehicle at intersection M – Safety Disabled

81.2

Kadence - Split & Offset Adjustment 84.8

Kadence - Offset & Sequence Adjustment 67.7

Kadence - Split & Sequence Adjustment 66.9

Kadence Offset, Split & Sequence Adjustment 84.8

223.6

84.5

77.1

97.9

78.8

77.1

86.0

371.5

107.3

75.7

203.2

78.4

75.7

128.4

646.2

166.5

70.1

256.9

91.8

70.1

Simulation Interval (seconds)

SynchroVISSIM

VISSIMVirtual D4

900 - 1800

63.4

125.0

1800 - 2700

67.5

2700 -3600 3600 - 4500

Kadence - Split Adjustment

110

Table 21. WB Left-turns - Average Delays per Vehicle at intersection M – Safety Disabled

83.7

Kadence - Split & Offset Adjustment 94.8

Kadence - Offset & Sequence Adjustment 104.1

Kadence - Split & Sequence Adjustment 88.0

Kadence Offset, Split & Sequence Adjustment 94.8

195.9

55.1

49.9

150.2

81.7

49.9

74.9

231.1

73.3

54.2

229.3

154.3

54.2

59.2

253.8

195.9

56.8

264.6

137.8

56.8

Simulation Interval (seconds)

SynchroVISSIM

VISSIMVirtual D4

900 - 1800

77.1

143.0

1800 - 2700

64.5

2700 - 3600 3600 - 4500

Kadence - Split Adjustment

6.4.5 Cases 5 through 7, Non-optimum phase sequence selection at intersection M These cases present different configuration of phase sequence selection at intersection M compared to the baseline model, which has lag left-turns at intersection M. In case 5, the eastbound left-turns was a lead left while the westbound left-turn was a lag left. In Case 6, the sequence of left-turns included a westbound leading left and a lag left for the eastbound. In case 7, both mainline left-turns are leading. Except for the phase sequence alterations all remaining parameters of the corridor layout, traffic volumes, and remaining signal timing and phasing remain the same as the baseline case. The objective of these test cases is to verify the functionality of the phase sequence tuning algorithm and determine if Kadence is able to select the most efficient phase sequence at intersection M. As shown in Table 22 and Table 23, the results obtained for these three cases were driven by a phase utilization for the left-turn movements below the Light Utilization threshold of 40-percent. Therefore, Kadence appeared to favor a lead left-turn phase sequence under these circumstances. Nevertheless, the results for all options were non-conclusive. Table 22. Average Delay per Vehicle at intersection M - No Adjustments

42.7

Case 5 VISSIM-Virtual D4 Lead-Lag 47.5

Case 6 VISSIM-Virtual D4 Lag-Lead 45.7

Case 7 VISSIM-Virtual D4 Lead-Lead 45.4

1800 - 2700

42.4

57.6

43.0

42.6

2700 - 3600

42.8

69.8

44.1

39.4

3600 - 4500

48.5

74.7

48.7

41.2

Simulation Interval (seconds)

Baseline Synchro-VISSIM Lag-Lag

900 - 1800

111

Table 23. Average Delays per Vehicle at Intersection M - Kadence Safety Disabled

42.7

Case 5 Safety Disabled Lead-Lag to LeadLead 45.1

Case 6 Safety Disabled Lag-Lead to LeadLead 44.0

1800 - 2700

42.4

48.5

41.6

42.6

2700 - 3600

42.8

49.6

39.8

39.4

3600 - 4500

48.5

40.7

41.4

41.2

Simulation Interval (seconds)

Baseline Synchro-VISSIM Lag-Lag

900 - 1800

Case 7 Safety Disabled Lead-Lead 45.4

Table 24 shows the queue length and total number of conflicts extracted from case 6. The results show that Kadence’ adjustments, when compared to the VISSIM-Virtual D4 scenario, reduced the total number of conflicts and also reduced the length of the queues. Table 24. Case 6 Results - Queue Lengths at intersection M SynchroVISSIM

VISSIMVirtual D4

Kadence Safety Disabled

Kadence Safety Enabled

900 - 1800

72

160

162

162

1800 - 2700

88

191

176

176

2700 - 3600

158

276

187

187

3600 – 4500

297

331

119

119

58.4

63.5

56.3

56.3

225/209

225/214

225/224

225/224

900 - 1800

108

189

144

144

1800 - 2700

91

351

159

159

2700 - 3600

102

605

128

128

3600 – 4500

82

561

114

114

34

52.7

47.3

47.3

205/206

205/175

205/211

205/211

Queue Length EBL

(feet)

Simulation Interval

MOE

Total Conflicts, EB Left-Turn (per 1,000 vehicles)

Queue Length WBL

(feet)

Simulation Interval

Vehicles, EB Left-Turn (input/simulated)

Total Conflicts, WB Left-Turn (per 1,000 vehicles) Vehicles, WB Left-Turn (input/simulated)

6.4.6 Modified baseline for cycle length adjustments A modified baseline scenario was developed to derive the test cases used to verify the cycle length tuning algorithm. In order to test the cycle adjustment algorithm, which selects cycle

112

lengths from pre-established scheduled patterns, another simulation model was created for different vehicular demand and, consequently, different optimized signal timing parameters. The additional simulation was designed to run a low cycle length (70 seconds) to serve a lower traffic demand. The low demand simulation was also created and optimized in Synchro for actuated and coordinated signal controllers. The low demand TOD pattern and volumes were later added to the original baseline VISSIM model to create a modified baseline scenario and a second timing pattern. Figure 65 shows the two patterns implemented in the modified baseline. The volume data in TOD Pattern #1 is the same as the demand used by the previous tests. Signal timing for this pattern was optimized in Synchro with a cycle length of 130 seconds. The second pattern was optimized in Synchro for a lower traffic demand and it is labeled as TOD Pattern #2, which runs a cycle length of 70 seconds. Table 25 presents the TOD schedule with the starting and ending times for each of the two patterns. Table 26 presents the actual demand interval schedule for the two patterns.

Figure 65. Illustration. Demand and signal timing information for scenario with nonoptimum offset at all intersections.

113

Table 25. TOD schedule for modified baseline case. TOD Pattern 1 - Low Cycle 2 - High Cycle

Start Time 9:00 AM 10:15 AM

End Time 10:15 AM 11:15 AM

Cycle Length 70 seconds 130 seconds

Table 26. Demand interval schedule for modified baseline case. Demand Level Low High

Start Time 9:00 AM 10:15 AM

End Time 10:15 AM 11:15 AM

The next four cases are derived from the above modified baseline scenario. In these cases, the start of the actual demand intervals and the start of the corresponding pre-planned TOD patterns are indifferent intentionally so that Kadence can be tested to determine if it can respond in realtime and take action to address the actual traffic demand conditions. 6.4.7 Case 8 - Higher Demand Starts Late This case is derived from the modified baseline scenario. Schedule for the TOD plans is per the plan listed in Table 26. The demand intervals are altered so that the low traffic demand can extend 30 minutes into the higher cycle TOD pattern. Figure 66 shows the pattern schedule and demand intervals.

Figure 66. Illustration. TOD schedule and demand intervals showing an extended period of low traffic demand. The objective of this test case is to verify the functionality of the cycle length tuning algorithm and determine if Kadence is capable of determining the change in traffic demand and, hence take action accordingly. Figure 67 confirms the results of Kadence, and notes that Kadence in fact implemented the higher cycle length later than planned and maintained the lower cycle length past its TOD schedule due to the lower traffic demand than expected.

114

Figure 67. Illustration. Kadence graph for cycle change in Case 8 Table 27 summarizes the average delay per vehicle generated by the adjustments caused by cycle length tuning algorithm. The results are not expected to show any significant improvement since this wasn’t the intent of the analysis. Nevertheless, the results show marginal differences in travel time between the modified baseline model and the Kadence signal timing settings for the maintaining the lower cycle TOD pattern. Table 27. Case 8 Results – Average Delays per Vehicle at each intersection VISSIMVirtual D4

Kadence Safety Disabled

900 - 1800

15.5

15.5

1800 - 2700

16.7

16.7

2700 - 3600

16.4

16.4

3600 – 4500

17.6

17.6

4500 - 5400

20.5

16.6

5400 - 6300

23.6

20.6

6300 - 7200

33.8

33.8

7200 - 8100 Total Intersection Conflicts (per 1,000 vehicles) Traffic Volume (input/simulated)

36.4

34.5

17.1

17.8

9175/8980

9175/8966

900 - 1800

21.7

21.7

1800 - 2700

22.2

22.2

2700 - 3600

21.2

21.2

3600 – 4500

22.4

22.4

M

Average Delay Per Vehicle (seconds)

Average Delay Per Vehicle (seconds)

Simulation Interval

W

MOE

Simulation Interval

Intersection

115

VISSIMVirtual D4

Kadence Safety Disabled

4500 - 5400

20.5

23.0

5400 - 6300

20.8

22.7

6300 - 7200

27.7

27.1

7200 - 8100 Total Intersection Conflicts (per 1,000 vehicles) Traffic Volume (input/simulated)

34.9

38.3

15.0

16.2

9240/9044

9240/9028

900 - 1800

13.7

13.7

1800 - 2700

14.7

14.7

2700 - 3600

15.9

15.9

3600 – 4500

13.6

13.6

4500 - 5400

20.6

15.9

5400 - 6300

20.2

16.2

6300 - 7200

28.9

29.1

7200 - 8100

32.4

37.4

15.3

16.4

9288/9021

9228/8998

MOE

Average Delay Per Vehicle (seconds) E

Simulation Interval

Intersection

Total Intersection Conflicts (per 1,000 vehicles) Traffic Volume (input/simulated)

6.4.8 Case 9 – High Demand Starts Early This case is derived from the modified baseline scenario. Schedule for the TOD plans is per the plan listed in Figure 68. The demand intervals are altered so that the high traffic demand can start 30 minutes earlier than anticipated, and therefore a higher cycle length will be implemented earlier than planned. Figure 68 shows a scheme of the pattern schedule and demand intervals.

Figure 68. Illustration. TOD schedule and demand intervals showing an abrupt increase in demand earlier than anticipated.

116

The objective of this test case is to verify the functionality of the cycle length tuning algorithm and determine if Kadence is capable of determining the change in traffic demand and, hence take action accordingly. Figure 69 confirms the results of Kadence, and notes that Kadence in fact implemented the higher cycle length earlier than planned and maintained the higher cycle length to the end of its TOD interval.

Figure 69. Illustration. Kadence graph of cycle change in Case 9 Table 28 summarizes the average delay per vehicle generated by the adjustments caused by cycle length tuning algorithm. Again, the results are not expected to show any significant improvement since this wasn’t the intent of the analysis. Nevertheless, the results show marginal differences in average delay per vehicle between the modified baseline model and the Kadence signal timing settings for the early implementation of the higher cycle TOD patterns. Table 28. Case 9 Results, average delays per vehicle at each intersection VISSIMVirtual D4

Kadence Safety Disabled

900 - 1800

17.1

16.4

1800 - 2700

17.8

17.3

2700 - 3600

25.1

25.4

3600 – 4500

27.9

33.4

4500 - 5400

36.4

45.3

5400 - 6300

36.7

41.4

6300 - 7200

34.1

35.7

7200 - 8100 Total Intersection Conflicts (per 1,000 vehicles)

37.6

44.9

20.5

15.9

W

MOE

Average Delay Per Vehicle (seconds)

Simulation Interval

Intersection

117

VISSIMVirtual D4

Kadence Safety Disabled

13145/13023

13145/13006

900 - 1800

22.1

23.0

1800 - 2700

21.6

21.5

2700 - 3600

32.6

33.2

3600 – 4500

35.4

33.4

4500 - 5400

31

31.9

5400 - 6300

34.5

33.3

6300 - 7200

34.7

33.7

7200 - 8100 Total Intersection Conflicts (per 1,000 vehicles) Traffic Volume (input/simulated)

35.3

35.5

20.4

17.1

13240/12131

13240/13130

900 - 1800

14.4

13.9

1800 - 2700

15.1

14.6

2700 - 3600

25.2

25.0

3600 – 4500

32.7

31.0

4500 - 5400

31.1

31.3

5400 - 6300

31.9

31.2

6300 - 7200

37.4

33.5

7200 - 8100

35.2

36.7

19.3

17.0

13223/13079

13223/13090

Intersection

MOE

Average Delay Per Vehicle (seconds) E

Simulation Interval

M

Average Delay Per Vehicle (seconds)

Simulation Interval

Traffic Volume (input/simulated)

Total Intersection Conflicts (per 1,000 vehicles) Traffic Volume (input/simulated)

6.4.9 Case 10- Low Demand Starts Late This case is derived from the modified baseline scenario. Schedule for the TOD plans is per the plan listed in Figure 70. The demand intervals, as listed in Table 29, are altered so that the high traffic demand can extend the higher cycle 30 minutes into the current TOD plan. Table 30 shows the pattern schedule and demand intervals.

118

Figure 70. Illustration. TOD schedule and demand intervals showing an extended period of high demand. Table 29. TOD schedule for modified baseline case. TOD Pattern 2 - High Cycle 1 - Low Cycle

Start Time 9:00 AM 10:15 AM

End Time 10:15 AM 11:15 AM

Cycle Length 70 seconds 130 seconds

Table 30. Demand interval schedule for modified baseline case. Demand Level High Low

Start Time 9:00 AM 10:45 AM

End Time 10:45 AM 11:15 AM

The objective of this test case is to verify the functionality of the cycle length tuning algorithm and determine if Kadence is capable of determining the change in traffic demand and, hence take action accordingly. Figure 71 confirms the results of Kadence, and notes that Kadence in fact maintained the higher cycle length 30 minutes past its scheduled TOD plan, and then selected a lower cycle length in the next TOD plan for lower traffic volumes. Table 31 summarizes the average delay per vehicle generated by the adjustments caused by cycle length tuning algorithm. Again, the results are not expected to show any significant improvement since this wasn’t the intent of the analysis. Nevertheless, the results show marginal differences in average delay per vehicle between the modified baseline model and the Kadence signal timing settings for the implementation of the higher cycle TOD pattern earlier than planned.

119

Figure 71. Illustration. Kadence graph for cycle change in Case 10 Table 31. Case 9 Results – Average Delays per Vehicle at each intersection VISSIMVirtual D4

Kadence Safety Disabled

900 - 1800

38.1

38.1

1800 - 2700

33.3

33.3

2700 - 3600

32.7

32.7

3600 – 4500

45.7

45.7

4500 - 5400

35.8

37.4

5400 - 6300

31.2

34.1

6300 - 7200

20.1

20.3

7200 - 8100 Total Conflicts Intersection (per 1,000 vehicles) Vehicles Intersection (input/simulated)

16.8

17.2

19.8

18.8

13145/13205

13145/13208

900 - 1800

35.2

35.2

1800 - 2700

36.0

36.0

2700 - 3600

32.8

32.8

3600 – 4500

30.9

30.9

4500 - 5400

39.7

38.2

5400 - 6300

35.7

32.4

6300 - 7200

25.8

25.0

7200 - 8100

21.3

21.6

20.0

15.5

M

Average Delay Per Vehicle (seconds)

Average Delay Per Vehicle (seconds)

Simulation Interval

W

MOE

Simulation Interval

Intersection

Total Conflicts Intersection

120

Intersection

VISSIMVirtual D4

Kadence Safety Disabled

13240/12239

13240/13330

900 - 1800

33.7

33.7

1800 - 2700

35.0

35.0

2700 - 3600

34.0

34.0

3600 – 4500

29.9

29.9

4500 - 5400

28.0

33.9

5400 - 6300

30.6

31.7

6300 - 7200

20.3

21.6

7200 - 8100

14.6

14.4

19.6

15.2

13230/13270

13223/13270

MOE (per 1,000 vehicles)

Average Delay Per Vehicle (seconds) E

Simulation Interval

Vehicles Intersection (input/simulated)

Total Conflicts Intersection (per 1,000 vehicles) Vehicles Intersection (input/simulated)

6.4.10 Case 11- Low Demand Starts Early This case derives from the modified baseline scenario. The corridor layout remains the same. The schedule for the TOD patterns is listed in . The demand intervals, however, are altered so the low demand interval starts 30 minutes before the low cycle TOD pattern is scheduled to start. Figure 72 shows the pattern schedule and demand intervals. Table 32 lists the demand interval schedule.

Figure 72. Illustration. Demand and signal timing information for scenario with nonoptimum offset at all intersections. Table 32. Demand interval schedule for modified baseline case. Demand Level High Low

Start Time 9:00 AM 9:45 AM

121

End Time 9:45 AM 11:15 AM

The objective of this test case is to verify the functionality of the cycle length tuning algorithm and determine if Kadence is capable of determining the change in traffic demand and, hence take action accordingly. Figure 73 confirms the results of Kadence, and notes that Kadence in fact identified the lower traffic demand and initiated the lower cycle length 30 minutes before its scheduled TOD plan; it also maintained the cycle length until the end of its scheduled TOD plan. Table 33 summarizes the average delay per vehicle generated by the adjustments caused by cycle length tuning algorithm. Again, the results are not expected to show any significant improvement since this wasn’t the intent of the analysis. Nevertheless, the results show marginal differences in average delay per vehicle between the modified baseline model and the Kadence signal timing settings for the implementation of the higher cycle TOD pattern earlier than planned.

Figure 73. Illustration. Kadence graph for cycle change in Case 11

122

Table 33. Case 11 Results – Average Delays per Vehicle at each intersection VISSIMVirtual D4

Kadence Safety Disabled

900 - 1800

34.3

34.3

1800 - 2700

32.7

32.7

2700 - 3600

25.8

25.8

3600 – 4500

23.1

18.6

4500 - 5400

19.3

17.1

5400 - 6300

17.2

17.1

6300 - 7200

16.9

16.9

7200 - 8100 Total Intersection Conflicts (per 1,000 vehicles) Traffic Volume (input/simulated)

17.6

17.8

16.0

14.2

9175/9307

9175/9301

900 - 1800

31.7

31.7

1800 - 2700

32.1

32.1

2700 - 3600

26.3

26.3

3600 – 4500

20.0

23.3

4500 - 5400

22.7

22.0

5400 - 6300

22.4

22.5

6300 - 7200

21.3

21.8

7200 - 8100 Total Intersection Conflicts (per 1,000 vehicles)

21.6

21.6

13.2

14.1

9240/9331

9240/9336

900 - 1800

31.5

31.5

1800 - 2700

37.0

37.0

2700 - 3600

24.9

24.9

3600 – 4500

22.2

17.0

4500 - 5400

19.4

13.8

5400 - 6300

17.8

16.5

6300 - 7200

16.0

16.4

7200 - 8100

15.2

14.7

14.4

15.0

9228/9349

9228/9314

W

Average Delay Per Vehicle (seconds) M

Simulation Interval

Average Delay Per Vehicle (seconds)

Simulation Interval

MOE

Intersection

Average Delay Per Vehicle (seconds) E

Simulation Interval

Traffic Volume (input/simulated)

Total Intersection Conflicts (per 1,000 vehicles) Traffic Volume (input/simulated)

123

Further testing was performed to verify the effects of combinational tuning parameters. Table 34 presents a summary of results for intersection M, whereby all combinations of tuning algorithms were tested. Although none of the results show any compelling reductions in delay, stops and queues but it was demonstrated the even for a single intersection, the Kadence solution was relatively close to the baseline optimal solution. 

Measures used in the analysis, as tabulated below, include the following:



Vehicle: average number of vehicles



Delay: average delay per vehicle in seconds



Stops: average number of stops per vehicle



Queue: average queue length in feet

124

No adjustment

0.2 0.4 1.2 0.3 0.4

11.5

5.6

77.8

17.4

4.4

287

29

31

297

47

EBT

EBR

WBL

WBT

WBT

45.6

11.5

5.6

77.8

17.4

4.4

VEHICLES

33

287

29

31

297

47

EBL

EBT

EBR

WBL

WBT

WBT

Movement

DELAY (s)

No adjustment

1

45.6

33

DELAY (s)

EBL

Movement

VEHICLES

0.4

0.3

1.2

0.4

0.2

1

0

20

20

0

50

40

0

20

20

0

50

40

0.3

17.4 0.4

1.2

77.8 4.4

0.4

0.2

11.5 5.6

1

0

20

20

0

50

40

Cycle Length + Split + Offset

47

297

31

29

287

33

45.6

Cycle Length Only

51

283

31

39

322

39

5.8

19.5

35.1

23.2

37.8

52.7

0.5

0.3

0.9

0.9

0.6

1.2

0

20

20

0

50

50

Cycle Length + Offset

10.3

21.7

58.5

3.0

10.2

40.4

0.5

0.4

1.2

0.2

0.2

1

0

20

20

0

50

40

Cycle Length + Offset + phase Sequence

54

298

34

29

274

38

52

10.8

23.2

57.7

18.2

36.2

73.8

125

301

33

35

316

43

0.6

0.4

1

1

0.6

1.2

0

30

30

0

50

40

Cycle Length + Split

15.1

29.0

49.9

28.0

35.8

57.8

0.7

0.5

0.9

0.9

0.6

1.1

0

20

20

0

50

40

Cycle Length + Split + Phase Sequence

53

311

28

37

325

39

53

315

28

37

327

39

16.3

29.6

53.2

26

34

61.9

0.8

0.5

1.1

0.8

0.6

1.2

0

20

20

0

60

40

Cycle Length + Phase Sequence

Table 34. Case 11 – Results for different combinations of adjustments – values for intersection M at simulation interval 2700 to 3600

STOPS STOPS

QUEUE (ft) QUEUE (ft)

VEHICLES VEHICLES

VEHICLES VEHICLES

DELAY (s) DELAY (s)

DELAY (s) DELAY (s)

STOPS STOPS

STOPS STOPS

QUEUE (ft) QUEUE (ft)

QUEUE (ft) QUEUE (ft)

VEHICLES VEHICLES

5.1

11.5

64.8

13.2

27.8

78.1

0.5

0.3

1

0.6

0.5

1.1

0

20

30

0

50

40

Cycle Length + Offset + Split + Phase Sequence

38

268

30

37

289

40

VEHICLES 51

284

31

39

322

39

VEHICLES

DELAY (s) DELAY (s)

DELAY (s) 5.8

19.4

35

23.2

37.8

52.8

DELAY (s)

STOPS STOPS

STOPS 0.5

0.3

0.9

0.9

0.6

1.2

STOPS

QUEUE (ft) QUEUE (ft)

QUEUE (ft) 0

20

20

0

50

50

QUEUE (ft)

6.5 SENSITIVITY TESTING OF SYSTEM PARAMETERS Sensitivity analysis was performed to verify the appropriateness of the various ranges in the Kadence system parameter data. Specifically, various tests were performed using different values for the parameters affecting offsets, splits, cycle lengths and phase utilization. Sensitivity analysis simulations were only performed for scenarios with the safety function disabled. Enabling the safety function wouldn’t add any value to this type of testing. Therefore, it was decided to isolate the safety function from the sensitivity analysis. This testing followed the same procedure used for the verification testing described above, and also used scenarios with non-optimum conditions so that Kadence can perform parameter tuning to develop more efficient solutions. The list of parameters, values and ranges tested, and the enabled tuning algorithm are shown in Table 35. Table 35. Sensitivity analysis parameters and tested values. System Parameter Max Offset Increment Max Offset Deviation Max Split Increment Max Split Deviation Heavy Utilization Light Utilization Dominate Alpha Cycle Decrease Saturation Threshold Cycle Increase Saturation Threshold Min Intersections to Jump Forward Min Intersections to Jump Back Run Cycle Interval Start Time Cycle Check End Time Cycle Check Cycle Increment

Values Tested 2 – 5 – 25 25 – 60 5 – 20 10 – 15 – 25 75 – 95 25 – 50 2–5 30 – 40 65 – 75 1–3 1–3 5 – 20 15 – 30 15 – 30 2–5

Tuning Algorithm Enabled Offset Offset Split Split Phase Sequence Phase Sequence Phase Sequence Cycle Length Cycle Length Cycle Length Cycle Length Cycle Length Cycle Length Cycle Length Cycle Length

Five test cases were selected for the sensitivity analysis. The list of testing cases selected to be analyzed and the corresponding tuning algorithm enabled is shown in Table 36.

126

Table 36. List of sensitivity analysis cases and respective tuning algorithm tested. Case Description Case 2 - Non-Optimum Offset Selection At All Intersections Case 4 - Over-Saturated Left-Turns (EB And WB) At Intersection M Case 7 - Non-Optimum Phase Sequence Selection At Intersection M Both Lead Case 8 - Higher Cycle Time-Of-Day Pattern Starts Late Case 10 - Lower Cycle Time-Of-Day Pattern Starts Late

Tuning Algorithm Offset Split Phase Sequence Cycle Length Cycle Length

6.5.1 Sensitivity Analysis for Offset Adjustment Parameters The test case with non-optimum offsets at all three intersections was selected for the sensitivity analysis of the offset tuning algorithm. Table 37 lists the parameter combination for each set of simulation. “Original” corresponds to the set of parameters used during the initial verification testing described earlier. Table 37. Offset parameter combinations for sensitivity analisys. Sensitivity Test Original test case A B C D E F

Max Offset Increment 10 5 2 10 10 5 25

Max Offset Deviation 45 45 45 25 60 25 60

Table 38 and Table 39 show the results of the offset parameter sensitivity test cases. Other than validating the proper functioning of the algorithm there were some notable changes in delays based on varying the offset increments and maximum deviation. Table 38. Mainline average delay per vehicle (safety disabled) – eastbound through Sensitivity Test VISSIMVirtual D4 Kadence Safety Disabled A B

Simulation Interval (seconds)

Parameters 900 - 1800

1800 - 2700

2700 - 3600

No adjustments made

58.0

62.7

70.4

Max Offset Incr = 10s Max Offset Dev = 45s

61.8

92.1

86.1

53.3

43.9

48.1

52.4

48.7

39.5

49.8

43.6

50.8

Max Offset Incr = 5s Max Offset Dev = 45s Max Offset Incr = 2s Max Offset Dev = 45s

127

3600 - 4500 64.4

Sensitivity Test C D E F

Simulation Interval (seconds)

Parameters

Max Offset Incr = 10s Max Offset Dev = 25s Max Offset Incr = 10s Max Offset Dev = 60s Max Offset Incr = 5s Max Offset Dev = 25s Max Offset Incr = 25s Max Offset Dev = 60s

900 - 1800

1800 - 2700

2700 - 3600

3600 - 4500

45.1

72.8

88.7

42.7

46.6

46.1

50.0

44.5

42.7

47.8

49.1

45.2

62.8

60.9

61.7

78.2

Table 39. Mainline average delay per vehicle (safety disabled) – westbound through Sensitivity Test VISSIMVirtual D4 Kadence Safety Disabled A B C D E F

Simulation Interval (seconds)

Parameters 900 - 1800

1800 - 2700

2700 - 3600

No adjustments made

26.0

50.2

61.9

Max Offset Incr = 10s Max Offset Dev = 45s

33.2

39.3

43.5

46.3

26.2

33.7

47.4

86.8

29.4

34.9

43.0

78.5

29.7

39.6

34.9

40.8

29.1

34.1

43.2

77.5

28.4

31.3

46.7

83.4

30.5

36.5

35.7

45.8

Max Offset Incr = 5s Max Offset Dev = 45s Max Offset Incr = 2s Max Offset Dev = 45s Max Offset Incr = 10s Max Offset Dev = 25s Max Offset Incr = 10s Max Offset Dev = 60s Max Offset Incr = 5s Max Offset Dev = 25s Max Offset Incr = 25s Max Offset Dev = 60s

3600 - 4500 89.7

The test case with over-saturated left-turns (eastbound and westbound) at intersection M was selected for the sensitivity analysis for the split tuning algorithm. Table 40 lists the parameter combination for each set of simulation. “Original test” corresponds to the set of parameters used during the verification testing process described earlier. Table 40. Split parameter combinations for sensitivity analisys Sensitivity Test Original test A B C

Max Split Increment 10 5 10 20

128

Max Split Deviation 45 45 10 45

Sensitivity Test D E

Max Split Increment 10 10

129

Max Split Deviation 25 15

Table 41 and Table 42 show the results of the split parameter sensitivity test cases. The range of parameters selected appears to make a notable difference on average delays. Table 41. Mainline average delay per vehicle (safety disabled) – eastbound left-turn Sensitivity Test VISSIMVirtual D4 Kadence Safety Disabled A B C D E

Simulation Interval (seconds)

Parameters

No adjustments made Max Split Incr = 10s Max Split Dev = 45s Max Split Incr = 5s Max Split Dev = 45s Max Split Incr = 10s Max Split Dev = 10s Max Split Incr = 20s Max Split Dev = 45s Max Split Incr = 10s Max Split Dev = 25s Max Split Incr = 10s Max Split Dev = 15s

900 - 1800

1800 - 2700

2700 - 3600

3600 - 4500

125.0

223.6

371.5

81.2

84.5

107.3

166.5

90.8

77.7

129.8

158.5

85.8

77.4

110.2

216.7

88.8

85.8

117.7

177.4

85.8

82.4

111.9

165.0

85.8

81.8

113.8

147.3

646.2

Table 42. Mainline average delay per vehicle (safety disabled) – westbound left-turn Sensitivity Test VISSIMVirtual D4 Kadence Safety Disabled A B C D E

Simulation Interval (seconds)

Parameters

No adjustments made Max Split Incr = 10s Max Split Dev = 45s Max Split Incr = 5s Max Split Dev = 45s Max Split Incr = 10s Max Split Dev = 10s Max Split Incr = 20s Max Split Dev = 45s Max Split Incr = 10s Max Split Dev = 25s Max Split Incr = 10s Max Split Dev = 15s

900 - 1800

1800 - 2700

2700 - 3600

3600 - 4500

143.0

195.9

231.1

253.8

83.7

55.1

73.3

195.9

112.4

101.8

207.1

215.6

98.8

64.9

159.4

180.7

96.2

68.7

103.8

168.0

98.8

59.3

80.3

248.2

98.8

57.9

148.7

209.2

The test case with non-optimum phase sequence selection at intersection M, in which both mainline left-turn movements lead, was selected for the sensitivity analysis for the phase sequence tuning algorithm. Table 43 lists the parameter combination for each set of simulation. “Original test” corresponds to the set of parameters used during the verification testing process described above. 130

Table 43. Phase Sequence parameter combinations for sensitivity analisys. Sensitivity Test Original A B C D E F

Heavy Utilization 85 75 95 85 85 85 85

Light utilization 40 40 40 50 25 40 40

Dominate Alpha 5 5 5 5 5 2 10

Table 44 shows Table 45 the results for phase sequence parameters sensitivity test cases. The findings were consistent and average delays did not change much under the range of variable parameter thresholds. Table 44. Mainline average delay per vehicle (safety disabled) – eastbound through Sensitivity Test VISSIMVirtual D4 Kadence Safety Disabled A

B

C

D

E

F

Simulation Interval (seconds)

Parameters

No adjustments made Heavy UT = 85% Light UT = 40% Dominate Alpha = 5 Heavy UT = 75% Light UT = 40% Dominate Alpha = 5 Heavy UT = 95% Light UT = 40% Dominate Alpha = 5 Heavy UT = 85% Light UT = 50% Dominate Alpha = 5 Heavy UT = 85% Light UT = 25% Dominate Alpha = 5 Heavy UT = 85% Light UT = 40% Dominate Alpha = 2 Heavy UT = 85% Light UT = 40% Dominate Alpha = 10

900 - 1800

1800 - 2700

2700 - 3600

3600 - 4500

46.5

38.1

31.4

39.8

38.6

36.7

35.0

46.5

38.1

31.4

37.0

46.5

38.1

31.4

37.0

46.5

38.1

31.4

37.0

46.5

38.1

31.4

37.0

46.5

38.1

31.4

37.0

46.5

38.1

31.4

37.0

37.0

The testing case with a lower cycle TOD pattern starting early was selected for the sensitivity analysis for the cycle length tuning algorithm. Table 45 lists the parameter combination for each set of simulation. “Original test” corresponds to the set of parameters used during the verification 131

testing process described above. The results of the simulations are presented in Table 46 and Table 47. The findings were consistent and average delays did not change much under the range of variable parameter thresholds for Decrease Sat Threshold, Increase Sat Threshold, Run Cycle Interval, Start Time Check, and End Time Check. However, the findings indicate that varying these parameters under various traffic conditions could make a noticeable change in average delays and other operation performance measures such as queues, stops, travel time.

Table 45. Cycle length parameter combinations for sensitivity analisys. Sensitivity Test

Original A B C D E F

Decrease Sat Threshold 50 50 50 50 50 50 35

Increase Sat Threshold 85 65 75 75 75 85 85

132

Run Cycle Interval 10 10 10 5 20 10 10

Start Time Check 60 60 60 60 60 15 30

End Time Check 60 60 60 60 60 30 30

Table 46. Mainline average delay per vehicle (safety disabled) – eastbound through Sensitivity Test VISSIMVirtual D4 Kadence Safety Disabled

A

B

C

D

E

F

Simulation Interval (seconds)

Parameters

No adjustments made Decrease Sat Threshold = 50% Increase Sat Threshold = 85% Run Cycle Interval = 10s Start Time Check = 60s End Time Check =60s Decrease Sat Threshold = 50% Increase Sat Threshold = 65% Run Cycle Interval = 10s Start Time Check = 60s End Time Check =60s Decrease Sat Threshold = 50% Increase Sat Threshold = 75% Run Cycle Interval = 10s Start Time Check = 60s End Time Check =60s Decrease Sat Threshold = 50% Increase Sat Threshold = 75% Run Cycle Interval = 5s Start Time Check = 60s End Time Check =60s Decrease Sat Threshold = 50% Increase Sat Threshold = 75% Run Cycle Interval = 20s Start Time Check = 60s End Time Check =60s Decrease Sat Threshold = 50% Increase Sat Threshold = 85% Run Cycle Interval = 10s Start Time Check = 15s End Time Check =30s Decrease Sat Threshold = 35% Increase Sat Threshold = 85% Run Cycle Interval = 10s Start Time Check = 30s End Time Check =30s

900 - 1800

1800 - 2700

2700 - 3600

3600 - 4500

32.7

21.5

15.4

6.9

32.7

21.5

16.4

29.5

31.0

22.3

14.9

29.2

31.0

22.3

14.9

29.2

31.0

22.3

14.9

29.2

31.0

22.3

14.9

29.2

31.0

22.3

14.7

9.5

31.0

22.3

14.7

9.5

133

Table 47. Mainline average delay per vehicle (safety disabled) – westbound through Sensitivity Test VISSIMVirtual D4 Kadence Safety Disabled

A

B

C

D

E

F

Simulation Interval (seconds)

Parameters

No adjustments made Decrease Sat Threshold = 50% Increase Sat Threshold = 85% Run Cycle Interval = 10s Start Time Check = 60s End Time Check =60s Decrease Sat Threshold = 50% Increase Sat Threshold = 65% Run Cycle Interval = 10s Start Time Check = 60s End Time Check =60s Decrease Sat Threshold = 50% Increase Sat Threshold = 75% Run Cycle Interval = 10s Start Time Check = 60s End Time Check =60s Decrease Sat Threshold = 50% Increase Sat Threshold = 75% Run Cycle Interval = 5s Start Time Check = 60s End Time Check =60s Decrease Sat Threshold = 50% Increase Sat Threshold = 75% Run Cycle Interval = 20s Start Time Check = 60s End Time Check =60s Decrease Sat Threshold = 50% Increase Sat Threshold = 85% Run Cycle Interval = 10s Start Time Check = 15s End Time Check =30s Decrease Sat Threshold = 35% Increase Sat Threshold = 85% Run Cycle Interval = 10s Start Time Check = 30s End Time Check =30s

4500 - 5400

5400 - 6300

6300 - 7200

7200 - 8100

46.5

38.1

31.4

37.0

25.7

29.5

30.1

28.1

28.0

29.5

30.1

27.1

27.7

29.1

29.5

27.3

27.7

29.1

29.5

27.3

27.7

29.1

29.5

27.3

27.7

29.1

29.5

27.3

26.4

28.9

29.4

27.7

134

6.6 SUMMARY Results of the various test cases and sensitivity analyses revealed the following: 

Verification of the offset tuning algorithm generated good results and demonstrated a reduction in travel time, delays and total number of conflicts when compared to nonoptimum signal timing plans. Although the results were not always consistent, the surrogate safety analysis demonstrated that Kadence is capable of generating efficient signal timing settings that are sensitive to surrogate safety measures, i.e. conflcits.



As shown in case 2, Kadence demonstrated that it can increase vehicle throughput by continuously tuning the offsets to maintain coordination and efficient progression. Likewise, in cases 3 and 4 Kadence demonstrated similar positive results of lesser delays and conflicts with the split tuning algorithm while servicing the increased volume of the left-tuning traffic.



In cases 5 through 7, the phase sequence tuning algorithm was demonstrated successfully although the benefits of reduced delays and queues were not very compelling. One of the reasons for the lack of favorable results was the fact that the majority of the left-turn phases had phase utilizations less than 70-percent. Under these cases, however, there was a modest reduction in the queue length and was reduced from 331 to 119 conflicts under the “Kadence Safety Disabled” and “Kadence Safety Enabled” alternatives, respectively.



Under cases 8 through 11, Kadence successfully demonstrated its ability to adjust the start and end time schedule for the TOD cycle length based on actual traffic demand, rather than perceived traffic demand. Several scenarios were tested and all showed favorable reduction in the total number of conflicts when the correct cycle length—one that responds to actual traffic demand—is implemented.



Kadence is able to adjust the transition length between two cycles based on a userdefined increments; adjusted by a one second increment or a 10-second increment. The incremental adjustment could be enabled with or without cycle switching, that early or late implementation of the next cycle.



It was unrealistic to base all testing, verification and findings of the analysis on a single 3-intersection arterial model and low to medium flow regimes. Nevertheless, the limited funding available under this project dictated that we focus our analysis and verification on the proof-of-concept rather than on testing all possible operational scenarios. Therefore, the results were convincing that Kadence has the capability to fine-tune signal timing and phase sequence in real-time to benefit both efficiency and safety.



The sensitivity analysis demonstrated that the Kadence results could change substantially based on the various parameter settings and configuration in the Kadence Parameter Menu. All thresholds can be sensitive to invoking the tuning algorithms, specifically thresholds that affect the phase utilization under near-saturated conditions. For example, choosing a wide range of split adjustment thresholds (see Table 33 and Table 34 ) could results in significant variance in average delay per vehicle. Under low phase utilization conditions, the Phase Utilization thresholds for phase sequence adjustment would not 135

have any significant effect on performance measures because it may not invoke the tuning algorithm. 

The combination of parameter settings for cycle length tuning, such as Decrease Saturation Thresholds, Increase Saturation Threshold, Run Cycle Interval, Start Time Check, and End Time Check (see Table 41 and Table 42) have noticeable impact on the average delay per vehicle, depending on the choices selected.



The tuning algorithms for cycle selection, offsets, splits, and phase sequence are promising and deserve further research to verify the operation of the various parametertuning algorithms under various traffic flow regimes and roadway configurations. The testing and verification of the four algorithms demonstrated in this report have successfully validated the concept of operations of all algorithms and the safety performance function. The sensitivity analysis also demonstrated that further research is still needed to develop better guidelines on how the ranges of the various signal timing parameters should be selected.



There are several individual and network-wide signal timing parameters and configuration choices the user should make to derive efficient and safer solutions than typical TOD plans would offer. Eventually, clear guidelines will be developed to assist the user in selecting these settings. Otherwise, there could be a significant amount of time and technical skills needed to fine tune a system deployment in the field.



Kadence performs best when all algorithms are enabled concurrently with cycle selection, phase sequence, splits, and offsets. Enabling one algorithm alone, while disabling the other algorithms, is not necessarily beneficial. Likewise, enabling one or multiple algorithms for a single intersection in a coordinated system isn’t beneficial either.



The safety performance function was validated successfully. Kadence demonstrated the relationship between optimized signal timing and surrogate safety measures. Nonetheless, surrogate safety measures were not always reduced with optimized signal timing plans.



The safety performance function developed in this project is only a performance index of surrogate safety measures associated with optimized or non-optimized signal timing settings. The SFP does not optimize the frequency of conflicts neither does it influence the optimization of signal timing settings.



The various types of conflicts were not segregated by type, therefore it is not clear how improved signal timing can effect rear-end, lane change and angle conflicts individually.



Unpublished traffic simulations scenarios demonstrated that Kadence, in its current state, doesn’t optimize signal timing settings for over-saturated traffic flow regimes.



Interested agencies should recognize that the adaptive algorithms developed under this project don’t substitute good engineering judgment and sound traffic engineering skills. In fact, implementing these adaptive strategies without good technical skills can possibly result in unsafe operations, poor signal timing, and worse traffic flow conditions than usually operated under TOD signal timing plans. 136

7 SIMULATION OF REAL-LIFE CASE SCENARIOS Subsequent to the testing of the Kadence tuning algorithms (refer to Sections 5 and 6 the research team decided to further validate the performance of the tuning algorithms by testing real-world networks in Software-in-Loop (SIL) simulation environment. A sample arterial and a grid network were selected for testing purposes and are discussed below. 7.1 ARTERIAL APPLICATION Figure 74 provides a snapshot of the Bell Road arterial network which was selected for testing the performance of Kadence adaptive system. This 3-mile segment of Bell Road is located in Surprise, Arizona and exemplifies the typical characteristics of a suburban arterial roadway network. The study network comprises of the following five signalized intersections along Bell Road (From West to East):     

Reems Road, Parkview Place, Bullard Avenue, Sun Village Parkway, and Litchfield Road

Figure 74. Illustration. Bell Road arterial network. The intersections are spaced ¼-mile to ½-mile apart. Portions of this roadway accommodate 70,000 vehicles per day. The arterial becomes oversaturated on several critical routes during the morning and afternoon peak periods due to surges in traffic, mainly caused by day-to-day variability in traffic conditions. Because of its relative location adjacent to major freeway connections to central Phoenix, there is a pattern of high commuter traffic eastbound in the AM

137

peak and westbound in the PM peak. In addition, the arterial experiences oversaturation when special event traffic demand is overlaid on the already heavy through flows. Each year during the months of February, March, and April the Baseball Training Facility at Surprise Stadium (located south of the intersection of Bell Road and Bullard Avenue) hosts two major league baseball (MLB) teams, the Texas Rangers and the Kansas City Royals, for spring training camps. Therefore, as a result of these special events the commuter traffic along Bell Road is increased substantially by the additional vehicles traveling to and from the Baseball Training Facility. The origin-destination patterns of event traffic overlaid with PM peak-hour traffic is illustrated i in Figure 75. The result of this overlap is a large increase in westbound left turning vehicles at Bullard which causes blocking and starvation for the through phases at the intersections upstream of Bullard to the East since the primary access to this part of the Valley is from the northbound or westbound route 101 loop freeway. When baseball games occur at night, the game generated traffic combined with the normal commuter traffic quickly causes excessive queuing and spillbacks in the arterial. Some oversaturation is also caused by the increase in traffic heading eastbound towards the game facility, but these backups are not nearly as notable as those in the westbound direction.

Stadium Figure 75. Illustration. Critical routes during game overlaid with peak flows Figure 76 provides a snapshot of typical weekday off-peak traffic volumes within the Bell Road network. The inset within the figure indicates turning volumes at Bullard Avenue during a typical event at the Stadium.

138

Figure 76. Illustration. Bell Road arterial network traffic volumes. The considerable variation in turning volumes under the off-peak conditions versus gamespecific conditions at Bullard Avenue are highlighted in Figure 76. 7.1.1 Simulation-in-Loop Model Development For the purposes of SIL testing a copy of the VISSIM database for the Bell Road arterial, prepared by Kimley-Horn And Associates, was utilized. The model was coded as part of an operational study and replicates the existing AM, Off-Peak, PM and event period traffic conditions. In order to establish a baseline condition for comparative analyses, the research team decided to generate a TOD schedule with optimized cycle/split/offset/sequence for all the different traffic volume conditions. The SYNCHRO traffic software was utilized to develop the optimized timing plans and subsequently imported into the Virtual D4 signal controllers. The simulation parameters were adjusted to run for 16,800 seconds encompassing all traffic patterns. The TOD pattern schedule assumed for simulation purposes is outlined in Table 48.

139

Table 48. Bell road TOD pattern schedule Begin Time Period 600 Sim Sec 4,200 Sim Sec 7,800 Sim Sec 11,400 Sim Sec 15,000 Sim Sec 18,600 Sim Sec*

TOD Pattern# Pattern 1 Pattern 2 Pattern 3 Pattern 4 Pattern 5 Pattern 6*

Cycle Length (seconds) 80 130 140 150 130 70*

*Although Pattern 6 falls outside the simulation time threshold, it was coded to verify adaptive algorithm performance

Figure 77 provides a snapshot of the Bell Road VISSIM network.

Figure 77. Illustration. VISSIM snap shot of the Bell Road arterial network. Figure 78 shows the maximum green times allotted for various phases under the base timing condition (Pattern 1).

140

Figure 78. Illustration. Bell Road arterial network- Pattern 1 green times. Three alternatives, as shown below, were developed for comparative analyses with the existing conditions, within the SIL simulation framework. Alternative A. Bell Road Network with Synchro optimized actuated-coordinated signal control Alternative B. Bell Road Network with Kadence Adaptive Control (SPF Disabled) Alternative C. Bell Road Network with Kadence Adaptive Control (SPF Enabled) The existing signal timing plans, represented by the existing field operations in the summaries of analyses, is based on timing pans that were recently optimized and deployed by the operating agency. Alternative A is expected to mimic the operation of a typical arterial with Synchro-optimized actuated-coordinated signal control under a rigid TOD pattern schedule. In this case the signal timing, although optimized, would not be able to adjust to varying traffic volumes and shifting patterns. Hence, it is anticipated to reproduce congested conditions similar to those typically observed during field operations. Alternative B allows the Kadence adaptive algorithms to tune the signal timing parameters without the safety performance function, i.e. SPF disabled. The algorithms discern traffic conditions based on the calculated phase utilization and captured flow from occupancy data and make adjustments to offsets, splits, phase sequence and cycle selection in real-time. Adjustments in signal timing parameters is expected to enhance efficiency and provide better performance measures than those derived in Alternative A. Alternative C also allows the Kadence adaptive algorithms to tune the signal timing parameters with the safety performance function, i.e. SPF enabled. Adjustments to splits, offsets, phase sequence and cycle selection are more deliberate under this alternative, and are only made when the predicted total number of conflicts, as the result of efficiency improvements, is lower than

141

the total number of conflicts for the existing signal timing conditions. Under this alternative, unless safety is improved the signal timing solution will be the same as those in Alternative B. 7.1.2 Simulation-in-the-Loop Assumptions Model geometrics, driver and vehicle behavior parameters, vehicle compositions, inputs and signal control devices and parameters remained consistent within the VISSIM simulation network utilized for analyzing each of the three alternatives described above. The one notable exception was the addition of an advance detector on the westbound left-turn lane at the intersection of Bell Road and Bullard Avenue for Alternatives B and C. This was done to enable data collection for the purposes of the Kadence algorithms. No pedestrian or bicycle volume was considered in the model. For each alternative, five simulation runs were performed with discrete random seeds to provide a statistical average. The simulation included a 10-minute seeding period, in which the network was populated with appropriate flow of vehicles. Data measurements were taken only after the seeding period, in 60-minute intervals per iteration. This approach ensured that the network was sufficiently loaded with vehicles prior to beginning collection of performance measures. The system parameter thresholds in Kadence were adjusted to make sure that the adaptive algorithms are sensitive to changes in the traffic patterns along Bell Road, and are able to adjust signal timing parameters to enhance efficiency (and safety, when SPF is enabled). A snapshot of the Kadence system parameters is shown in Figure 79 for reference.

Figure 79. Illustration. Kadence system parameters for the Bell Road arterial network. 142

7.1.3 Performance Measures Performance measures were derived from the SIL simulation testing at the network, corridor and intersection levels. Network-wide performance measures included average delay, total delay, number of stops, and total travel time. Corridor performance measures included travel time data collected along selected segments of Bell Road. Intersection level performance measures consisted of movement delay. Conflict data was derived in SSAM by processing vehicle trajectory information from the VISSIM traffic simulation model. Based on the description of the existing conditions, it is evident that the critical operating conditions occur at the time of overlap between events at the baseball stadium and commuter peak hours. Further, the intersection of Bell Road and Bullard Avenue becomes the key junction, (as it is the primary access point to the baseball stadium) whose operation impacts the entire network. Hence, a comparison of performance measures at this location would provide suitable validation of the Kadence adaptive algorithms. Table 49 summarizes the safety and performance measures derived at this location. Table 49. Bell Road/Bullard Avenue – 15-Minute MOEs (6000 s to 6900 s) Alternative A Alternative B Alternative C MOE (Synchro Kadence with Kadence with Optimized) SPF Enabled SPF Enabled Average Intersection Delay – LOS (seconds) Total Conflicts (per 1,000 vehicles) Vehicles (input)(simulated)

24.3 (C)

22.3 (C)

23.0 (C)

12.2

9.8

8.3

(1244)(1225)

(1244)(1248)

(1244)(1236)

Although the intersection delay measure shows a modest improvement, a review of the total conflicts as well as the vehicles processed in the simulation model also show a modest decrease in the total number of conflicts. The total conflicts are reduced by 20% and 32% under Alternatives B and C, respectively, as compared to Alternative A. The results illustrate that when the Kadence adaptive algorithms are enabled, both safety and efficiency are improved under the prevailing traffic conditions. Table 50 and Table 51 summarize the efficiency and surrogate safety performance measures derived at Bell Road and Reems Road, and Bell Road and Parkview Place, respectively.

143

Table 50. Bell Road/Litchfield Road - 15-Minute MOEs (6000 s to 6900 s) MOE Average Intersection Delay – LOS (seconds) Total Conflicts (per 1,000 vehicles) Vehicles (input/simulated)

Alternative A

Alternative B

Alternative C

55.9 (E)

47.4 (D)

54.5 (D)

17.0

16.2

13.3

1210/1187

1210/1264

1210/1225

Table 51. Bell Road/Reems Road - 15-Minute MOEs (6000 s to 6900 s) MOE Average Intersection Delay – LOS (seconds) Total Conflicts (per 1,000 vehicles) Vehicles (input/simulated)

Alternative A

Alternative B

Alternative C

31.9 (C)

29.3 (C)

28.6 (C)

17.6

15.6

15.3

1267/1320

1267/1292

1267/1306

Figure 80 and Figure 81 illustrate the offset and split adjustments performed during the SIL simulation testing of Alternative B at the intersection of Bell Road and Bullard Avenue.

144

Figure 80. Illustration. Bell Road/Bullard Avenue-Kadence offset adjustments.

Figure 81. Illustration. Bell Road/Bullard Avenue- Kadence split adjustments. 145

Kadence’s adjustment to signal timing parameters, under Alternatives A and B, include real-time split tuning to Phase 5 (westbound left-turn movement) to accommodate the increased traffic demand for traffic destined to the baseball stadium. Figure 82 through Figure 86 illustrate the performance measures derived on a network-wide basis during the SIL simulation analysis. The total number of conflicts is derived from SSAM. The “Existing Field Operations” represents the optimized TOD signal plan as implemented by the operating agency.

Figure 82. Illustration. Bell Road arterial network- Average delay time comparison. When timings are adjusted by the Kadence adaptive system (SPF disabled), the average delay time per-vehicle is shown to be 14% lower than the average delay for the existing field operations as well as the Synchro optimized solution. When SPF is enabled, the results show approximately 6% of reduction in the average delay.

146

Figure 83. Illustration. Bell Road arterial network-Total delay time comparison. Total delay shown above presents a 12.7% reduction over existing and Synchro-optimized conditions when SPF is disabled in Kadence. With the SPF enabled, the reduction in total delay is approximately 5%.

147

Figure 84. Illustration. Bell Road arterial network: Comparison of number of stops. The total number of stops shown above presents a 6% reduction over existing conditions when SPF is disabled in Kadence. The reduction is improved to 12% when compared with the Synchro optimized solution. Therefore, the surrogate safety performance associated with signal timing of both alternatives of Kadence was better than those associated with signal timing of the existing field operations and the Synchro-optimized solution. However, when comparing both alternatives of Kadence, the number of stops increased by 6% when SPF was enabled. This anomaly is not unexpected since it was stated earlier that the average error in the safety performance function was 17%, and also the SPF is not an optimizer of surrogate safety measures.

148

Figure 85. Illustration. Bell Road arterial network- comparison of total travel time. As shown in Figure 85, the total travel time within the network is reduced by 2% over existing conditions when SPF is disabled in Kadence. In summary, the following observations are noted: 

Traffic flow conditions analyzed under all alternatives were below congested flow regimes.



Kadence is able to recognize surges in traffic demand and, hence vary the signal timing parameters in real-time to improve efficiency. Efficiency improvement was consistent, but improvement in the surrogate safety measures were not consistent.



Under the three alternatives tested, Kadence did not make any changes in the phase sequence and cycle length selection. This is not necessarily a shortcoming of Kadence but may be explained by the phase utilization and phase split selection criteria and the fact that none of the alternatives experienced congested conditions.

149

7.2 GRID APPLICATION Unlike an arterial network which typifies a suburban setting, a grid system is characteristic of urban city centers. Intersections tend to be more closely spaced and traffic control may or may not comprise of signals. As such traffic flow patterns are more staggered. For the purposes of the testing a grid network, the research team selected a 3 by 3 grid in downtown DC, comprising of nine signalized intersections. Figure 86 illustrates the limits of this grid network.

Figure 86. Illustration. DC Grid network The grid network consists of 12th Street NW, 13th Street NW and 14th Street NW (which run in a North-South direction) intersecting with E Street NW, F Street NW and G Street NW (which run in an East-West direction). 7.2.1 Simulation-in-Loop Model Development The grid network was developed by District Department of Transportation (DDOT) as part of the Mall Area and Vicinity Simulation (MAVS, Washington DC) in 2010. Under this project, traffic simulation models of the downtown core area of DC and 6 major corridors were modeled to study various operational scenarios, including peak-hour activity and outbound traffic patterns following major events on the Mall such as the Fourth of July (Operation Fast Forward). The project provided DDOT with a planning tool to quantitatively support decision making processes for congestion mitigation strategies for special events as well as for daily operations. Figure 87 provides a snapshot of the VISSIM network.

150

Figure 87. Illustration. DC Grid network- VISSIM model. Intersection turning movement counts for this grid was collected in 15-minute intervals during the AM, Mid-Day, PM and Off-Peak time periods. Figure 88 illustrates the typical weekday peak hour traffic volumes traversing the grid network.

Figure 88. Illustration. DC Grid network turning movement volumes.

151

Existing signal timing data, including the TOD pattern schedule, was obtained from DDOT as part of this project. The grid network, extracted from the MAVS project, was updated with the traffic volume and TOD pattern data collected. These patterns are outlined in Table 52. Table 52. DC Grid SIL- TOD pattern schedule Begin Time Period

TOD Pattern#

Cycle Length (seconds)

0 Sim Sec 1,800 Sim Sec 7,200 Sim Sec 14,400 Sim Sec

Pattern 1 Pattern 2 Pattern 3 Pattern 4

80 100 100 100

It should be noted that although Patterns 2, 3 and 4 run the same cycle length of 100 seconds, the offsets and splits are different for each pattern. Figure 89 provides an indication of the green times typically allotted to phases at intersections.

Figure 89. Illustration. DC Grid network- Pattern 1 green times Similar to the arterial network, three alternatives were developed for comparative analyses within the SIL simulation framework. 152

A. B. C.

Grid Network with an optimized (Synchro) Actuated-Coordinated Signal Control Grid Network with Kadence Adaptive Control (SPF Disabled) Grid Network with Kadence Adaptive Control (SPF Enabled)

Alternative A is expected to mimic the operation of a typical arterial with Synchro optimized actuated-coordinated signal control under a rigid TOD pattern schedule. In this case the signal timing, although optimized, would not be able to adjust to varying traffic volumes and shifting patterns. Hence, it is anticipated to reproduce congested conditions similar to those typically observed during field operations. Alternative B allows the Kadence adaptive algorithms to tune the signal timing parameters without the safety performance function. The algorithms discern traffic conditions based on the calculated phase utilization and captured flow from occupancy data and make adjustments to offsets, splits, phase sequence and cycle selection in real-time. Adjustments in signal timing parameters is expected to enhance efficiency and provide better performance measures than those derived in Alternative A. Alternative C also allows the Kadence adaptive algorithms to tune the signal timing parameters with the safety performance function. Adjustments to splits, offsets, phase sequence and cycle selection are more deliberate under this alternative, and are only made when the predicted total number of conflicts, as the result of efficiency improvements, is lower than the total number of conflicts for the existing signal timing conditions. Under this alternative, unless safety is improved the signal timing solution will be the same as those in Alternative B. 7.2.2 Simulation-in-the-Loop Assumptions All model geometrics, driver and vehicle behavior parameters, vehicle compositions, inputs and signal control devices and parameters remained consistent within the VISSIM simulation network utilized for analyzing each of the three alternatives described above. No pedestrian or bicycle volume was considered in the model. For each alternative, five simulation runs were performed with discrete random seeds to provide a statistical average. The simulation included a 10 minute seeding period, in which the network was populated with appropriate flow of vehicles. Data measurements were taken only after the seeding period, in 60-minute intervals per iteration. This approach ensured that the network was sufficiently loaded with vehicles prior to beginning collection of performance measures. Each iteration was run for a time period of 21,600 seconds. The system parameter thresholds in Kadence were adjusted to make sure that the adaptive algorithms are particularly sensitive to changes in the traffic patterns along 12th Street NW and 14th Street NW, and adjust timing parameters to enhance efficiency (and safety, when SPF enabled) in a timely fashion. This was done based on evaluation of traffic volume and flow patterns. A snapshot of the Kadence system parameters is provided in Figure 90.

153

Figure 90. Illustration. DC Grid network- Kadence system parameters 7.2.3 Performance Measures Performance measures were derived from the SIL simulation testing at the network, corridor and intersection levels. Network-wide performance measures included average delay, total delay, number of stops, total travel time. Conflict data was obtained from SSAM by processing vehicle trajectory files obtained from VISSIM. Corridor performance measures included travel time data collected along selected segments of the Grid network. Intersection level performance measures consisted of movement delay. Based on the traffic patterns and volumes, the intersection of 14th Street, E Street and Pennsylvania Avenue NW is considered to be critical. Table 53 below provides a snapshot of the safety and performance measures derived at this location. Table 53. 15-Minute MOEs for 14th Street, E Street and Pennsylvania Avenue NW. MOE

Alternative A

Alternative B

Alternative C

Average Intersection Delay – LOS (seconds)

21.4 (C)

12.9 (B)

20.7 (C)

Total Conflicts (per 1,000 vehicles)

14.3

19.1

17.7

753/669

753/670

753/667

Vehicles (input/simulated)

154

The intersection delay shows a substantial improvement under Alternative B; however, the total conflicts increased even though the total throughput did not change from Alternative A. Figure 91 and Figure 92 illustrate the changes made in Kadence to the splits and offsets in response to the variation in traffic demand.

Figure 91. Illustration. 14th Street, E Street and Pennsylvania Avenue NW: Kadence offset adjustments

155

Figure 92 . Illustration. 14th St, E St and Pennsylvania Ave NW- Kadence split adjustments These figures demonstrate that Kadence is able to change splits and offsets in real-time to address fluctuation and variation in traffic demand for different turning movements. Figure 93 through Figure 96 illustrate the performance measures derived on a network-wide basis during the SIL simulation analysis. The data is juxtaposed with the total number of conflicts (obtained from SSAM) for each alternative on a secondary axis.

156

Figure 93. Illustration. Grid network: Average delay comparison

157

Figure 94. Illustration. Grid network- Total delay time comparison.

158

Figure 95. Illustration. Grid network- Comparison of numbers of stops.

159

Figure 96. Illustration. Grid Network- Comparison of total travel time. Overall, the results from the SIL simulation analyses of each of the three alternatives were not conclusive. A comparison among various alternatives tested in the SIL simulation at critical intersections indicates a noticeable improvement in performance, while safety shows either no improvement or a marginal increase in the total number of conflicts. 7.3 SUMMARY The Kadence adaptive algorithms were tested within the SIL simulation framework on two real world networks: a suburban arterial network which facilities commuter traffic and special event traffic, and an urban grid network with typical weekday traffic conditions. Both test case were limited to moderate traffic flow regimes and geometric conditions but don’t necessarily represent the wide variety of traffic and geometric conditions that exist in the real world. Kadence was tested with and without the SPF. Multiple iterations of the simulation were executed and performance measures were summarized on a network, corridor and intersection basis. Salient observations from the SIL simulation are summarized below:  

Kadence was tested successfully for both the arterial and the grid network. Kadence is cognizant of variations in traffic demand and responds by making suitable adjustments in real-time to splits and offsets.

160



Albeit modest improvements, Kadence benefits under the grid network were not conclusive. However, improvement in efficiencies was noticeable under the arterial test case.

161

162

8 FUTURE RESEARCH AND RECOMMENDED ENHANCEMENTS The following features are suggested enhancements for further development in Kadence or any model that would replicate the algorithms of Kadence. 1. “Groundhog Day” elimination. The current system tunes offsets and splits starting from a pre-configured timing plan. If these settings are inappropriate and the traffic flows are repeatable by day and time, Kadence will re-learn to adjust the settings each time it runs the same pattern. This effect is similar to the Bill Murray film “Groundhog Day” where he is forced to re-live the same day and events over and over. Future consideration should be given to calculate a new “starting point” for each TOD signal plan, as well as the most appropriate schedule start point for each timing plan. 2. Handling of oversaturated conditions. The current algorithms principles are all based on assumptions of under-saturated operation. While the system can be effective in reducing the duration of oversaturation, or delaying the start of oversaturated operation, in certain complex situations, the decisions made by Kadence may be counterproductive. This is true of every adaptive control system that is available today. Principles from NCHRP 03-90 research developed by Dr. Gettman will be incorporated into Kadence in the future to handle the special situations where downstream congestion creates upstream oversaturation or when phases are starved based on conflicting queues. These features require measurement of the queue length using the upstream detectors and estimation of the oversaturation levels. Gating links and other features related to the research in NCHRP 03-90 will be implemented. 3. Suggesting TOD signal timing plans and schedules. The same algorithms that can be applied in real-time can also be applied to longer periods of time (entire pattern durations) to determine suggested changes to baseline timing settings. This feature can provide more conservative agencies with benefits if they do not wish to adjust timings in real-time but rather want the system to suggest changes but not implement them until they can be reviewed by the engineering staff. 4. Free mode TOD plan selection: Enable full Kadence functionality when a TOD schedule includes a period where one or more signals are operated in “free” mode. 5. Cycle length selection: Enable Kadence to select a cycle length that exist in a library of TOD timing plans and not limit it to select only from the next or the previous cycles. 6. Segregation of conflicts by type: Currently all conflicts are combined in the safety performance function into a total number of conflicts. It is suggested to segregate the conflicts by type to determine how improved signal timing can affect the type of conflicts. It is further suggested that weighting factors should be used with each type of conflicts to signify the severity of the conflict; that could change the results of the surrogate safety performance index significantly.. 7. Performance safety function. Rather than using an absolute number for the SPF, it is suggested that a conflict rate would be more useful to account for the increased or reduced

163

throughput as a result of the signal timing optimization. This will require retraining the neural network. 8. Field pilot demonstration project. Kadence could be demonstrated in the field at a cost not exceeding $250K. A few agencies are interested to demonstrate the concept and validate the operation of the algorithms. However, financial assistance will be required. Nonetheless, a field demonstration could assist in validating the concepts of the efficiency model, the safety performance function, will assist in gauging the effort needed to fine-tune the system and controller parameters, and also to develop guidelines for field deployments.

164

REFERENCES 1. Ye, K.O. (1998). Orthogonal Column Latin Hybercubes and Their applications in Computer Experiments. Journal of the American Statistical Association. 2. Gettman, D., Pu, L., Sayed, T., and Shelby, S. (2008). Surrogate Safety Assessment Model and Validation: Final Report, Report No. FHWA-HRT-08-051, Federal Highway Administration, Washington, DC. 3. Trafficware. (2012). SIMTRAFFIC 7. Obtained from: www.trafficware.com. Site last accessed 4. PTV AG. (2008). VISSIM® User Manual, Version 5.1, Release 7, Karlsruhe, Germany. 5. Gettman, D., Shelby, S., Pu, L., and Joshi, R. (2008). Surrogate Safety Assessment Model (SSAM)—Software User Manual, Report No. FHWA-HRT-08-050, Federal Highway Administration, Washington, DC. 6. Gettman, D., et. al, (2006), ASCLITE Overview, FHWA-HRT-06-083, Washington, D.C. 7. Gettman, D., et. al., (2012), NCHRP Report No. 03-90, Final Report, Operations of Traffic Signal Systems in Oversaturated Conditions, to be published in National Academy of Sciences, Washington, D.C. 8. Sabra, Z.A., et. al, (2010), “Balancing Safety and Capacity in an Adaptive Signal Control System- Phase 1, Report No. FHWA-HRT-10-038, Federal Highway Administration, Washington, D.C. 9. T. Gordon, et. al., (2012), “Site-Based Video System Design and Development, SHAP2 Report, S2-S09-RW-1, National Academy of Science, Washington, D.C.

165

166

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

Appendix: Enhancing  Safety  and  Capacity  in  an  Adaptive  Signal  Control  System    -­  Phase  2   508 Compliance Report Figure 1. Illustration. Test network for experimental design. This illustration shows a simple network that was used as the basis for the test cases for experimental design. The network is comprised of three intersections. The mainline lays along the east-west alignment while the side streets are oriented north-south. The intersections are labeled “W”, “M”, and “E”, for “West”, “Middle”, and “East”, respectively. Lanes highlighted in green are used for Surrogate Safety Assessment Methodology (SSAM) analysis. Starting from the top left and moving clockwise in the figure, the highlighted links are labeled “Link M-W”, Link “N-M”, Link “E-M”, Link M-E”, and Link “W-M”. Also highlighted in green are the links to southbound and northbound approaches at intersection “M”, labeled “Link N-M” and “Link S-M”, respectively. Figure 2. Illustration. Variable and fixed arrival volumes. This illustration shows the variable and fixed arrival volumes. It shows the simple network described in Figure 1 as a base. Arrows are placed at all entry links of the network to indicate the volume arrival input. The arrows are colored blue in case of variable arrival volume, and red in case of fixed arrival volume. Starting from the top left and moving clockwise, the arrows are colored as follows: red, blue, red, blue, red, blue, red, and blue. Figure 3. Illustration. Right turn routes. This illustration shows an example of right turn routes. It shows the simple network described in Figure 1 as a base. Lanes used for SSAM analysis are highlighted in green and as follows: link N-M, link E-M, link S-M, and link W-M. The illustration shows five possible routes from the northbound approach of intersection W. Of all routes shown, three are making a right turn onto the mainline to contribute to the performance evaluation and are shown in red and blue routes indicating fixed traffic volume proportion and variable proportion, respectively. The two remaining routes are the route going through the intersection and the route left-turning onto the mainline. The routes are indicated as follows: left turn and exit (red), straight through and exit (red), right turn and continue straight until exit (blue), right turn and then left turn at M and exit (blue), right turn and then left turn at E and exit (red). Figure 4. Illustration. Through routes. This illustration shows an example of through routes. It shows the simple network described in Figure 1 as a base. Lanes used for SSAM analysis are highlighted in green and as follows: link N-M, link E-M, link S-M, and link W-M. The illustration shows six possible routes from a through entrance at W that are shown in red and blue routes indicating fixed traffic volume proportion and variable proportion, respectively. The routes are indicated as follows: left turn and exit (red), straight through and then left turn at M (blue), straight through and then left turn at E (blue), straight through and exit (blue), straight through and then right turn at E and exit (red), straight through and then right turn at M and exit (blue). Figure 5. Illustration. Left turn routes. This illustration shows an example of left turn routes. It shows the simple network described in Figure 1 as a base. Lanes used for

508 Compliance Report

A - 1  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

SSAM analysis are highlighted in green and as follows: link N-M, link E-M, link S-M, and link W-M. The illustration shows five possible routes from the southbound approach of intersection W. of all routes shown, three are making a left turn onto the mainline to contribute to the performance evaluation and are shown in red and blue routes indicating fixed traffic volume proportion and variable proportion, respectively. The two remaining routes are the route going through the intersection and the route left-turning onto the mainline. The routes are indicated as follows: right turn and exit (red), straight through and exit (red), left turn and then right turn at M and exit (blue), left turn and then right turn at E and exit (red), left turn and exit straight through (blue). Figure 6. Illustration. Side street approach routes. This illustration shows an example of side street approach routes. It shows the simple network described in Figure 1 as a base. Lanes used for SSAM analysis are highlighted in green and as follows: link N-M, link E-M, link S-M, and link W-M. The illustration shows five possible routes from the side street, northbound approach at M. The routes are shown in red and blue indicating fixed traffic volume proportion and variable proportion, respectively. There are five possible routes from the northbound side street approach routes at M: left turn and exit (blue), left turn and then right turn at W to exit (red), straight through and exit (red), right turn and then left turn at E to exit (red), right turn and exit (blue). Figure 7. Illustration. Link lengths modified in experimental design. This illustration indicates the links lengths varied for the experimental design. It shows the simple network described in Figure 1 as a base. Lanes used for SSAM analysis are highlighted in green and as follows: link N-M, link E-M, link S-M, and link W-M. Blue double sided arrows shown between intersections W and M, and between intersections M and E indicate the variable distance. Figure 8. Illustration. Splits adjusted in the experimental design. This illustration shows an example of the splits that were varied for the experimental design. It shows the simple network described in Figure 1 as a base. Lanes used for SSAM analysis are highlighted in green and as follows: link N-M, link E-M, link S-M, and link W-M. Webster fixed time splits are symbolized by a pie chart with 4 slices. Variable splits are symbolized by a pie chart with eight slices. Intersections W and E show the pie charts symbol for Webster fixed time splits. Intersection M shows a pie chart symbol for variable splits. Figure 9. Illustration. Offsets adjusted in the experimental design. This illustration shows an example of offsets adjusted in the experimental design. It shows the simple network described in Figure 1 as a base. Lanes used for SSAM analysis are highlighted in green and as follows: link N-M, link E-M, link S-M, and link W-M. The W and E intersections show a blue element that represents variable offsets. The M intersection shows a red element that represents a fixed offset. Figure 10. Illustration. Basic concept of a “neuron” in a neural network. This illustration shows the signal inputs displaced vertically and named xo, x1, x2, …, xp, and are symbolized by a small circle. Each small circle is connected by an arrow pointing to a

508 Compliance Report

A - 2  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

bigger circle which are called synaptic weights, indicated by wk0, wk1, wk2, …, wkp. Each connection between the input signals and the synaptic weights is connected by an arrow to the summation junction circle (∑). The summation is connected to the activation function box by an arrow. A threshold element is also connected to the activation function box. An arrow coming out of the activation box indicates the process output, called vk. Figure 11. Illustration. Sigmoid activation function. The sigmoid activation function is a common activation function for neural networks. The illustration shows a sigmoid function plotted in a xy graph. The sigmoid function will only produce positive numbers between 0 and 1. The sigmoid activation function is most useful for training data that is also between 0 and 1. Because the sigmoid activation function has a derivative, it can be used with gradient descent based training methods. Figure 12. Illustration. Illustration of the training process. This is a flowchart that shows the supervised learning process where a neural network matches the inputs to the outputs as closely as possible by adjusting the weights at each of the neurons since the correct outputs are known for the test cases. The input feature leads to the neural net. The neural net plus the target feature lead to the result. If there is an error vector, it is fed back to a supervised learning algorithm with a weight threshold adjustment, which then leads back to the neural net and begins the cycle again. Figure 13. Illustration. Neural network training tool in MATLAB. This is a screenshot of the neural network toolbox in the MATLAB software. The first portion shows the neural network flow diagram starting with the input which leads to a first layer, then a second layer and results in an output. The second section of the toolbox indicates the algorithms in use for training, performance and data division. The third section shows the training progress, and indicates epoch rated from 0 to 1000, time, performance, gradient from 1.00 to 1.00e-06, validation checks 0-6. The final section of the toolbox includes buttons for performance, training state, confusion, and receiver operating characteristic. Also included is a sliding scale for plot interval measured in epochs. Figure 14. Illustration. Stop bar and advance detection. This illustration shows a schematics of the detection equipment used by adaptive systems at an intersection. Phase utilization detectors are shown as long blue rectangles, aligned flush with the stop bar at each lane. Advance detectors, also called flow profiling detectors, are shown as blue squares and located 150 to 400 feet before the stop bar. Two notes are indicated in the graphic, “Detectors at stop-bar for split, cycle and sequence tuning” and “Set-back zones for offset tuning. Figure 15. Illustration. Flow chart of cycle time tuning algorithm. This illustration shows a flow chart that represents the cycle time tuning algorithm used in Kadence. First, the average phase utilization (PU) for designated critical phases is calculated. If the PU is greater than the critical level to increase the cycle, the safety impact is then calculated. If the safety is degraded, the current cycle would be retained. If the cycle is not degraded, the cycle should be increased or decreased. If the PU was not greater than the critical

508 Compliance Report

A - 3  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

level to increase the cycle and the PU was less than the critical level to decrease the cycle, the safety impact would be calculated and the cycle would be retained, increased or decreased based on whether or not the safety was degraded. If the PU is not less than the critical level to decrease the cycle, the current cycle would be retained. Figure 16. Illustration. Concept of implementing the next cycle time earlier or later than scheduled. This illustrations shows a graph of the time frame in which the decision is made to either implement the cycle length that is next in the time-of-day (TOD) schedule earlier or later than originally planned. The y axis represents the cycle time, in seconds, while the x axis represents the time of day. A blue line represents the cycle lengths. Three TOD plans are indicated across the top of the graph as “Plan 1”, which is set to run from 6:00 AM to 11:00 AM, “Plan 2”, which is set to run from 11:00 AM to 4:00 PM, and “Plan 3”, which is set to run from 4 pm to 9 PM. A note in the graph states that at 9:30 AM the cycle selection algorithm can start considering the next cycle. Another note in the graph states that at 1:45 PM, if Plan 1’s cycle length is still being enforced, than the cycle length will drop and Plan 2’s cycle length will be implemented. Figure 17. Illustration. Cycle selection process. This illustration shows the flowchart that outlines the algorithm described in Figure 16. It considers implementing the next pattern early in the TOD schedule if it meets a phase utilization threshold for a given number of minutes before the next pattern scheduled time. First the average phase utilization, or PU, is calculated. If the PU is not lesser than the defined threshold at a given amount of minutes before the next TOD is scheduled to change, and it is not greater than the threshold at a given amount of minutes after the TOD is scheduled to change, the next TOD pattern is implemented. If the PU is greater than the defined threshold at a given amount of minutes before and after the TOD is scheduled to change, the process moves forward to the next step. If the PU is greater than the critical level to increase the cycle length and the cycle length in the next scheduled TOD is higher than the current cycle, the safety impact is calculated. If the safety is not degraded, the next scheduled TOD plan is implements. If safety is degraded, the current TOD pattern in retained. If the PU is not greater that the critical level to increase the cycle length and the PU is not less than the critical level to decrease the cycle length, the current TOD plan is retained. If the PU is less than the critical level to decrease the cycle length and thecycle length in the next TOD plan is lower, the safety impact is calculated. If the safety is degraded, the current TOD plan is retained. If the safety is not degraded, the next TOD plan is implemented. Figure 18. Illustration. Typical flow profile detector locations on coordinated approaches. This illustration shows the locations of the flow profile detectors used for offset tuning. A simplified network schematics shows the upstream intersection as a “1”, and the downstream intersection as a “2”. A more detailed graphic of intersection “2” indicating the flow profile detectors is shown in this illustration .Detection equipment is shown for Intersection “2”. Flow profiling detectors are shown as blue squares, and are located 200-300 feet beforebehind intersection “2” stop bar.

508 Compliance Report

A - 4  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

Figure 19. Illustration. Example of traffic volume count and occupancy data from a typical advance detector. This figure shows a graphic of the cyclic flow profiles for six consecutive cycles. The graphic shows volume and occupancy recorded during the course of each cycle. In this figure, occupancy is represented by a blue bar while volume is represented by a red bar. Figure 20. Illustration. Example of phase timing for each of the last several cycles. This figure illustrates an example of historical phase timing observed over several cycles at an intersection. The number and colors in the timeline corresponds to the active phase interval (green, yellow, and red) in the cycle, displayed by each ring. Figure 21. Illustration. Example of average cyclic volume and occupancy profiles. The graphic in this figure shows measurements of volume, occupancy and phase green probability during the course of one cycle. Volume is represented in red bars; occupancy is represented in blue bars; and phase green is represented in green bars. The bars vary in height according to magnitude of the measurement they represent. The phase green probability is most concentrated during the period in which volume and occupancy are concentrated. Figure 22. Illustration. Offset adjustment algorithm flow chart. This flow chart uses the safety performance function as the performance calculator. For all intersections, calculate the percentage of arrivals on green for all phases. Next combine total percentage of arrivals for inbound and outbound approaches. Next, calculate the total percentage of arrivals for alternative offsets and choose highest percentage of arrivals. Next calculate the safety impact. If the safety is degraded, retain the current offsets. If the safety is not degraded, implement offset adjustment and repeats cycle. Figure 23. Illustration. Ring diagram with barriers. This diagram shows the typical layout for the ring diagram with barriers. The diagram, representative of a table, contains seven columns. The first column, referred to the barrier, is labeled “b”. The second column contains two rows, noted as phases “1” and “5”. Column three contains two rows, noted as phases “2” and “6”. Column four, referred to a barrier, is labeled “A”. The fifth column contains two rows, noted as phases “3” and “7”. Column six contains two rows, noted as phases “4” and “8”. The seventh column, referred to a barrier, is labeled “b”. Figure 24. Illustration. Detector layout. This illustration shows a single intersection with blue rectangles in the lanes noting phase utilization detector locations. Located over 100’ back from the stop bars, flow profile detector locations are noted as blue squares. Notes include “detectors at stop-bar for split, cycle and sequence tuning” and “set-back zones for offset tuning (>100’ from stop bar)”. Figure 25. Illustration. Measuring phase utilization for coordinated-actuated controllers. This illustration shows two horizontal bar graphs with two bars each. The bars represent the number of seconds of detector presence, shown in black, and the phase green time, shown in green. The top bar graph indicates the values for a fixed-time control. The bottom bar graph indicates the values for a coordinated-actuated control.

508 Compliance Report

A - 5  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

Below the bar graphs, the text details indicate the time measured in actual green time, split green time and available green time. Figure 26. Illustration. Utilization of phases before and after split adjustment. This illustration shows two vertical bar graphs. The first bar graph depicts the phase utilization obtained by pre-adjustment settings. The second bar graph depicts the phase utilization obtained by post-adjustment settings. Both graphs represent the phase utilization for each one of the eight phases, in percentage. A red text box in the first graph indicates that the pre-adjustment utilization for phase three reached 100%. After the adjustment, the phase utilization is reduced to 72%, which is indicated in the second graph by another red text box. Figure 27. Illustration. Flow chart of the split optimization process including safety analysis. This flow chart shows that for all intersection, the phase utilization for all phases is calculated. Then, the splits are re-allocated to minimize the maximum phase utilization. Next, the safety impact is calculated. If the safety is degraded, the current cycle is retained. If the safety is not degraded, new splits are implemented. Figure 28. Illustration. Adjustment algorithms flow chart. This flow chart depicts the sequence in which the adjustments are performed by the adaptive software. First, the algorithm reallocates the splits for all intersections. Next, offsets for all intersections are calculated. Next, phase sequence changes for all intersections are determined. If the phase sequence changes, offsets must be recalculated and the process, repeated. If phase sequence is not changed, the algorithm proceeds to adjust the cycle time. If the cycle is modified, the splits must be reallocated and the process repeats. If the cycle is not modified, the download of all settings is stopped. Figure 29. Illustration. Kadence Software-in-the-Loop Simulation. This is a graphical representation of the software-in-the-loop (SIL) process. The process starts with the VISSIM traffic simulation model that sends vehicular detector information to the Virtual D4 traffic signal controller. The Virtual D4 controller shares the data with Kadence that stores the information is a SQL database. Kadence processes the information, calculates the new settings, and finally sends them back to the Virtual D4 controller to be used by VISSIM. Figure 30. Illustration. Kadence field deployment schematics. This illustration shows a graphical diagram of Kadence field deployment. Real-time traffic and signal data are obtained by the Traffic Management Center, or TMC, in communication with the street controller equipment. The TMC shares the field data with Kadence. Kadence processes the existing data and determines the new improved settings. The TMC sends the new settings back to the controller to be applied in the field. Figure 31. Illustration. Kadence System Parameters. This is a screenshot of the application window of Kadence’s System Parameters settings. The parameters are listed in sections that correspond to each of the adjustment algorithms, such as cycle length, phase sequence, phase treatment (currently not in use), offset, and split. In a general

508 Compliance Report

A - 6  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

form, to the right side of each parameter name is placed a white text box that allows the user to entry or modify its value directly on the screen. To the right side of the text box is placed the range of values for each parameter. There are also three check boxes that enable plan switching cycle adjustment, incremental cycle adjustment, and the safety performance function. The user can also save the parameters as the default settings for future new projects. Figure 32. Illustration. Controller configuration parameters. This is a screenshot of the Kadence’s controller configuration window. It shows the various user-selectable fields. Fields include name of intersection, enable and disable button for adaptive logic at the controller, checked boxes indicate the desired adaptive logic algorithms, file path, controller number, controller description, timing, biasing, cycle, phase sequence boxes, and deviation. Figure 33. Illustration. Detector configuration parameters. This is a screenshot of Kadence’s detector configuration window. This application window shows a table with thirtheen columns representing the detector parameter settings for a given intersection. The rows in the table represent each one of the detector equipment. The first column is the detector number. Second column is the description. Third column is the signal phase called by the detector. Columns 4 and 5 are used to indicate if the detector is being used to collect phase utilization (stop bar detector) or for flow profiling (advanced detector). Columns 6 and 7 are the upstream and downstream controllers, respectively. Column 8 is the distance between the sop bar and the controller, and it is different than zero for advanced detectors only. Column 9 is the detector physical length. Column 10 is the free flow speed on the link where the detector is placed. Column 11 is the time to flow and indicates the time, in seconds, to travel from the detector location to the stop bar, at freeflow speed. Column 12 is the green time extension and it is not being used at this time. Finally, column 13 is the “sec to break point” parameter which indicates the time, in seconds, that it takes to safely stop by the stop bar, if the driver starts breaking when passing by the detector. Figure 34. Illustration. Link definition parameters. This is a screenshot of the Link parameters. This window is divided into three fields. Field 1 is a table in which the rows represent the “from” intersection and the columns represent the “to” intersection. Field 2 shows the scheme of the link. Field 3 is populated with data that corresponds to the phase number of the selected link direction, the distance between the two selected intersections, and the free flow speed on the link. Figure 35. Illustration. Phase configuration parameters. This is a screenshot of Kadence’s configuration window for a given intersection. This application window shows a table with fifteen columns representing the configuration parameters. The rows represent each of the phases for the given intersection traffic signal. The first column is the phase number. Columns 2 through 5 are the minimum green time, yellow, all red and maximum split times, respectively. Column 6 is Max2. Columns 7 through 9 are check boxes for minimum recall, maximum recall and pedestrian recall. Columns 11 through 15

508 Compliance Report

A - 7  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

correspond to walk time, pedestrian clearance, pedestrian delay, added initial, maximum initial and minimum gap. Figure 36. Illustration. Ring-barrier sequence. This screenshot shows the ring sequence for a given controller that is imported from the controller database. The screenshot displays the data in a table that cannot be modified. The table contains three columns: ring, phase sequence, and ring sequence showing the phases and the barriers. In this particular example, there are two rings displayed. Figure 37. Illustration. Pattern definition parameters. This screenshot shows the time-of-day pattern configurations for a given intersection. The application displays the parameters in two tables. The first table contains the data from the TOD patterns that are imported from the field for each controller. There are 22 columns containing the following data: pattern number, cycle time, offset time, alternate sequence, coordination mode, coordinated phase, and splits for phases 1 through phase 16. The second table displays the unit coordination patterns. This table contains six columns including: transition mode, maximum split mode, force off mode, offset reference, default coordination mode, and default alternate sequence. Figure 38. Illustration. TOD plan parameters. This screenshot shows the patterns schedule which is imported directly from the controller database. It contains the period of the day when each signal timing pattern begins and ends. The screenshot displays three sections: day plan configuration, timebase ASC action configuration, and unit timebase parameters. The day plan configuration displays a table with four columns: day plan, plan event, plan time, and pattern. The timebase ASVaction configuration is a grey out area. The unit timebase parameter displays a table with three columns: parameters, parameter value, and description. Figure 39. Illustration. SIL Data Flow. This illustration is a flow chart that represents the software-in-the-loop data flow process. The chart shows the open exchange communication and input to the virtual D4 controller interface by both the traffic simulation software (VISSIM) and the virtual controller database. Figure 40. Illustration. VISSIM Study Network. This screenshot shows the study network represented in the VISSIM interface. The study network model consists of a three-intersection corridor. Figure 41. Illustration. Typical Kadence log file illustration. Shows a screenshot of a portion of the log file showing the start of the adaptive logic process and detector information of a given controller. The screenshot shows the log file divided in eight fields delimited by red boxes: PC time, controller, detectors connected to controller, TOD, phase number in Ring 1, interval of the phase in Ring 1, phase number in Ring 2, and the interval of the phase in Ring 2. Figure 42. Illustration. Typical output graphs. This is a screenshot of the plan data display window and illustrates the output graphs generated in Kadence. The illustration

508 Compliance Report

A - 8  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

highlights nine fields: “Select Log File” button, a dropdown menu of the controllers with the generate button, offset adjustment graph, cycle adjustment graph, sequence adjustment graph, parameter adjustment graphs or the safety performance graphs selection, split adjustment for Ring 1 phase, and the split adjustment for Ring 2 graph. Figure 43. Illustration. Offset adjustment log (with SPF enabled). This screenshot is a sample of the offset adjustment’s log from Kadence. In this example, four fields are highlighted: offsets, MOE, total number of conflicts, and the offset selected by the algorithm. Figure 44. Illustration. Offset adjustment and safety prediction output graphs. This screenshot represents the graphical output of the offset tuning. The topmost image, a bar graph of offset adjustment, shows the offset changes along the simulation run. Simulation time is represented across the horizontal axis and controller offset (in seconds) is represented on the vertical axis. The bottom graph of offset safety shows the number of predicted and current estimation of conflicts at the instant of the given adjustment. The simulation time is represented across the bottom and the number of conflicts is represented on the left. Figure 45. Illustration. Sequence adjustment log. This screenshot is the output information of the decision process of the phase sequence adjustment. Six fields are highlighted: phase number, PU, left-turn sequence, result of the sequence, PU percentage as compared to the thresholds, and which approach or left-turn phase has the higher PU. Figure 46. Illustration. Sequence adjustment graph. This is the graphical representation of the phase sequence adjustment as it is presented by Kadence’s user interface. The vertical axis represents the sequence that phases are called. Sequences numbered 1 through 4 correspond to ring 1; sequences 5 through 8 refer to ring 2. Figure 47. Illustration. Split adjustment log. This screenshot of the split adjustment log file highlights eight fields: ring-barrier divider, phase number, actual split, PU value, action of the algorithm based on the phases; utilization, and the minimum and maximum bounds for the given phase split and its actual value at that instant. Figure 48. Illustration. Split adjustment Ring 1. This is a graphical representation of the trade-offs among phases for ring 1 and ring 2, respectively, during the split adjustment process. Simulation time is represented across the horizontal axis and phase time (seconds) is represented on the vertical axis. Phases 1 through 4 are represented by different colors, all on the same graph. Split safety worse, if any, is noted by a red dot. Figure 49. Illustration. Split adjustment Ring 2. This is a graphical representation of the trade-offs among phases for ring 1 and ring 2, respectively, during the split adjustment process. Simulation time is represented across the horizontal axis and phase time (seconds) is represented on the vertical axis. Phases 5 through 8 are represented by different colors, all on the same graph.

508 Compliance Report

A - 9  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

Figure 50. Illustration. Split surrogate safety measures prediction graph. This graph depicts the instant during the simulation when the splits are adjusted. Simulation time is represented across the horizontal axis and number of conflicts is represented on the vertical axis. Both predicted and current conflicts are noted in red and blue, respectively. Figure 51. Illustration. Log illustration of plan switching from higher to lower cycle. This screenshot illustrates the cycle adjustment indicated in the log file. It highlight three fields: PU value, actual planned cycle length, and the next plan and cycle length selected by Kadence. Figure 52. Illustration. Graphical illustration of plan switching from higher to lower cycle. This graph shows the graphical representation of the cycle length adjustment using the TOD plan-switching algorithm. Simulation time is represented along the horizontal axis and cycle length (seconds) is represented on the vertical axis. In this case, the actual cycle length running is 150 seconds and Kadence switches it to the next plan that has a cycle of 130 seconds. Figure 53. Illustration. Graphical illustration of incremental adjustment from higher to lower cycle. This graph shows the adjusted cycle length by increments of time that are added or withdrawn from the actual cycle to optimize the performance and safety (if enabled). Simulation time is represented along the horizontal axis and cycle length (seconds) is represented on the vertical axis. Figure 54. Illustration. Baseline arterial configuration. This illustration shows a schematic of the baseline arterial model. The baseline scenario consists of a corridor with three signalized intersections on a major street. The intersections are named as intersection M (middle), intersection W (west), and intersection E (east). The distance from W to M is noted as 1,200 feet and the distance from M to E is noted as 1,000 feet. Figure 55. Lane configuration and detection location. This illustration shows an intersection with stop bar detectors noted as purple rectangles and the advanced detectors noted as purple squares set upstream from the stop bars. Figure 56. Illustration. Demand and signal timing for the baseline scenario. This illustration shows the offsets, approach volumes, cycle length, and phase sequence for the intersections labeled W, M, and E. Figure 57. Illustration. Node locations. This screenshot from VISSIM shows the typical three-intersection layout. Each intersection is highlighted with a red box and noted as node 1, node 2, and node 3 for intersections W, M, and E, respectively. Figure 58. Illustration. Travel time segments. This screenshot from VISSIM shows the typical three-intersection layout. The travel time segments are highlighted as a red arrow noting westbound and a blue arrow noting eastbound across the three intersections along the main road.

508 Compliance Report

A - 10  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

Figure 59. Illustration. Demand and signal timing for Case 1. This illustration presents a non-optimum offset selection at intersection M. The corridor layout, traffic volumes, and all remaining signal timing and phasing parameters are the same as the baseline case. This graphic shows schematics of the traffic characteristics and depicts in a shadowed area the parameter that is modified. In this case, the offset at intersection M is set to 88 seconds. The baseline optimum offset is 63 seconds. All three of the intersection’s main street numbers include 350, 650, and 205 leading to the intersection. The W side street numbers from the south include 265, 2100, and 200 and from the north 425, 1980, and 185 leading to the intersection. The M side street numbers from the south include 280, 2150, and 225 and from the north 315, 2035, and 205 leading to the intersection. The E side street numbers from the south include 320, 2200, and 185 and from the north include 300, 2000, and 195 leading to the intersection. Figure 60. Illustration. Demand and signal timings for Case 2. This illustration presents non-optimum offset selection at all three intersections. In this case, the offsets at intersections W, M and E are set to 53 seconds, 88 seconds, and 101 seconds, respectively. The baseline optimum offsets are 78 seconds, 63 seconds, and 76 seconds, respectively. All three of the intersection’s main street numbers include 350, 650, and 205 leading to the intersection. The W side street numbers from the south include 265, 2100, and 200 and from the north 425, 1980, and 185 leading to the intersection. The M side street numbers from the south include 280, 2150, and 225 and from the north 315, 2035, and 205 leading to the intersection. The E side street numbers from the south include 320, 2200, and 185 and from the north include 300, 2000, and 195 leading to the intersection. Figure 61. Illustration. Demand and signal timings for Case 3. This illustration shows a schematic of the traffic characteristics for this test scenario, and highlights the eastbound left turn volume at intersection M, which is increased from 200 vph to 340 vph. All three of the intersection’s main street numbers include 350, 650, and 205 leading to the intersection. The W side street numbers from the south include 265, 2100, and 200 and from the north 425, 1980, and 185 leading to the intersection. The M side street numbers from the south include 280, 2150, and 340 (this number is highlighted in red) and from the north 315, 2035, and 205 leading to the intersection. The E side street numbers from the south include 320, 2200, and 185 and from the north include 300, 2000, and 195 leading to the intersection. Figure 62. Split adjustments for Case 3. This graph shows the split adjustment for ring 2 (phases five through eight) and phases are represented by different colors, all on the same graph. Simulation time is represented across the horizontal axis and phase time (seconds) is represented on the vertical axis. Kadence adjusted the phase splits favorably and reallocated time from phase 6 (eastbound through) to phase 5 (eastbound left). Figure 63. Illustration. Demand and signal timings for Case 4. This illustration shows a schematic of the traffic characteristics for this test case and highlights the east and westbound left turn volumes at intersection M. The left-turn demand at intersection M is increased from 200 vph to 340 vph for eastbound and from 185 vph to 310 vph for the

508 Compliance Report

A - 11  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

westbound, respectively. All three of the intersection’s main street numbers include 350, 650, and 205 leading to the intersection. The W side street numbers from the south include 265, 2100, and 200 and from the north 425, 1980, and 185 leading to the intersection. The M side street numbers from the south include 280, 2150, and 340 (this number is highlighted in red) and from the north 315, 2035, and 310 (this number is highlighted in red) leading to the intersection. The E side street numbers from the south include 320, 2200, and 185 and from the north include 300, 2000, and 195 leading to the intersection. Figure 64. Illustration. Split adjustments for Case 4. This illustration contains two graphs, one for the spit adjustment: ring 1 and the other for the split adjustment: ring 2. For each graph, simulation time is represented across the horizontal axis and phase time (seconds) is represented on the vertical axis. For ring 1, phases 1 through 4 are represented by different colors, all on the same graph. Split safety worse, if any, is noted by a red dot. For ring 2, phases 5 through 8 are represented by different colors, all on the same graph. Figure 65. Illustration. Demand and signal timing information for scenario with non-optimum offset at all intersections. This illustration shows two patterns implemented in the modified baseline. The first pattern, labeled as TOD Pattern #1, is the same as the demand used by the previous tests. Signal timing for this pattern was optimized in Synchro with a cycle length of 130 seconds. The second pattern was optimized in Synchro for a lower traffic demand and it is labeled as TOD Pattern #2, which runs a cycle length of 70 seconds. For TOD pattern #2, all three of the intersection’s main street numbers include 105, 325, and 75 leading to the intersection. The W side street numbers from the south include 135, 1150, and 100 and from the north 215, 890, and 95 leading to the intersection. The M side street numbers from the south include 140, 1075, and 115 and from the north 160, 1020, and 100 leading to the intersection. The E side street numbers from the south include 160, 1000, and 95 and from the north include 150, 1100, and 100 leading to the intersection. Below the two TOD patterns, a graphic representing time from 9:00 am to 11:15 am is shown with arrows representing duration. Low cycle length (70 s) TOD and low demand durations are shown from 9:00 am to 10:15 am and high cycle length (130 s TOD) and high demand are shown from 10:15 am to 11:15 am. Figure 66. Illustration. TOD schedule and demand intervals showing an extended period of low traffic demand. This graphic contains a timeline with a duration from 9:00 am to 11:15 am. Below the timeline are arrows representing duration intervals. Low cycle length (70 s) TOD is shown from 9:00 am to 10:15 am and high cycle length (130 s TOD) is shown from 10:15 am to 11:15 am. The low demand duration is shown from 9:00 am to 10:45 am and the high demand duration is shown from 10:45 am to 11:15 am. Figure 67. Illustration. Kadence graph for cycle change in Case 8. This graph for cycle adjustment shows simulation time across the horizontal axis and cycle length (seconds) on the vertical axis. Cycle is represented by the color green and is steady at 70s from 9:1:14 until 10:36:14, when it rises to 130s and remains until 11:15.

508 Compliance Report

A - 12  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

Figure 68. Illustration. TOD schedule and demand intervals showing an abrupt increase in demand earlier than anticipated. This graphic contains a timeline with a duration of 9:00 am to 11:15 am. Below the timeline are arrows representing duration intervals. Low cycle length (70 s) TOD is shown from 9:00 am to 10:15 am and high cycle length (130 s TOD) is shown from 10:15 am to 11:15 am. The low demand duration is shown from 9:00 am to 9:45 am and the high demand duration is shown from 9:45 am to 11:15 am. Figure 69. Illustration. Kadence graph of cycle change in Case 9. This graph for cycle adjustment shows simulation time across the horizontal axis and cycle length (seconds) on the vertical axis. Cycle is represented by the color green and is steady at 70s from 9:1:14 until 9:58:14, when it rises to 130s and remains until 11:15. Figure 70. Illustration. TOD schedule and demand intervals showing an extended period of high demand. This graphic contains a timeline with a duration from 9:00 am to 11:15 am. Below the timeline are arrows representing duration intervals. Low cycle length (70 s) TOD is shown from 9:00 am to 10:15 am and high cycle length (130 s TOD) is shown from 10:15 am to 11:15 am. The low demand duration is shown from 9:00 am to 10:45 am and the high demand duration is shown from 10:45 am to 11:15 am. Figure 71. Illustration. Kadence graph for cycle change in Case 10. This graph for cycle adjustment shows simulation time across the horizontal axis and cycle length (seconds) on the vertical axis. Cycle is represented by the color green and is steady at 130s from 9:1:14 until 10:43:0, when it falls to 70s and remains until 11:15. Figure 72. Illustration. Demand and signal timing information for scenario with non-optimum offset at all intersections. This graphic contains a timeline with a duration from 9:00 am to 11:15 am. Below the timeline are arrows representing duration intervals. High cycle length (130 s) TOD is shown from 9:00 am to 10:15 am and low cycle length (70 s TOD) is shown from 10:15 am to 11:15 am. The high demand duration is shown from 9:00 am to 9:45 am and the low demand duration is shown from 9:45 am to 11:15 am. Figure 73. Illustration. Kadence graph for cycle change in Case 11. This graph for cycle adjustment shows simulation time across the horizontal axis and cycle length (seconds) on the vertical axis. Cycle is represented by the color green and is steady at 130s from 9:1:14 until 9:58:14, when it falls to 70s and remains until 11:15. Figure 74. Illustration. Bell Road arterial network. This is an aerial photo snapshot of the 3-mile segment of the Bell Road arterial network. The study network comprises of the following five signalized intersections along Bell Road (From West to East): Reems Road, Parkview Place, Bullard Avenue, Sun Village Parkway, and Litchfield Road and are noted with a traffic signal graphic. The study roadways are highlighted in red. A major area just south of Bell Road is highlighted in yellow and labeled as the “Baseball Training Facility”.

508 Compliance Report

A - 13  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

Figure 75. Illustration. Critical routes during game overlaid with peak flows©. This is an aerial photo portion of the Bell Road arterial network. (Bell Road at Bullard Avenue). Blue arrows, indicating traffic flow patterns are overlaid on the aerial. The westbound arrows are thicker, indicating heavier traffic flow than the eastbound traffic. Five red dots are shown, representing the areas of excessive queuing and spillbacks in the arterial from the left-hand turns during baseball games. Figure 76. Illustration. Bell Road arterial network traffic volumes. This diagram illustrates turning volumes under the off-peak conditions versus game-specific conditions at Bullard Avenue. Figure 77. Illustration. VISSIM snap shot of the Bell Road arterial network. This image is a screenshot of the Bell Road VISSIM network. Figure 78. Illustration. Bell Road arterial network- Pattern 1 green times. This diagram shows the maximum green times allotted for various phases under the base timing condition. Figure 79. Illustration. Kadence system parameters for the Bell Road arterial network. This is a screenshot of the system parameters. Main categories include: cycle time parameters, miscellaneous, phase sequence parameters, protected permitted parameters, offset parameters, and split parameters. Figure 80. Illustration. Bell Road/Bullard Avenue-Kadence offset adjustments. Simulation time is represented across the horizontal axis and controller offset (seconds) is represented on the vertical axis. The offset is charted in green. Offset safety worse is noted by as a red line. Figure 81. Illustration. Bell Road/Bullard Avenue- Kadence split adjustments. This is a graphical representation of the trade-offs among phases for ring 2 during the split adjustment process. Simulation time is represented across the horizontal axis and phase time (seconds) is represented on the vertical axis. Phases 5 through 8 are represented by different colors, all on the same graph. Figure 82. Illustration. Bell Road arterial network- Average delay time comparison. This bar graph illustrates the performance measures derived on a network-wide basis during the SIL simulation analysis. The total number of conflicts is derived from SSAM. The “Existing Field Operations” represents the optimized TOD signal plan as implemented by the operating agency. Four alternatives are charted along the horizontal axis, delay time (seconds) is measured on the left side of the vertical axis and total conflicts are measured on the right side of the vertical axis. Also charted on this graph are total conflicts, represented in red. The Alternatives are as follows from left to right: existing field operations (79), actuated-coordinated signal control (80), Kadence SPF disabled (68), and Kadence SPF enabled (74). The total conflicts are charted as follows: existing field operations (1104), actuated-coordinated signal control (1219), Kadence SPF disabled (1121), and Kadence SPF enabled (1179).

508 Compliance Report

A - 14  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

Figure 83. Illustration. Bell Road arterial network-Total delay time comparison. This bar graph illustrates the performance measures derived on a network-wide basis during the SIL simulation analysis. The total number of conflicts is derived from SSAM. The “Existing Field Operations” represents the optimized TOD signal plan as implemented by the operating agency.Four alternatives are charted along the horizontal axis, delay time (seconds) is measured on the left side of the vertical axis and total conflicts are measured on the right side of the vertical axis. Also charted on this graph are total conflicts, represented in red. The Alternatives are as follows from left to right: existing field operations (612), actuated-coordinated signal control (626), Kadence SPF disabled (534), and Kadence SPF enabled (584). The total conflicts are charted as follows: existing field operations (1104), actuated-coordinated signal control (1219), Kadence SPF disabled (1121), and Kadence SPF enabled (1179). Figure 84. Illustration. Bell Road arterial network: Comparison of number of stops. This bar graph illustrates the performance measures derived on a network-wide basis during the SIL simulation analysis. The total number of conflicts is derived from SSAM. The “Existing Field Operations” represents the optimized TOD signal plan as implemented by the operating agency. Four alternatives are charted along the horizontal axis, delay time (seconds) is measured on the left side of the vertical axis and total conflicts are measured on the right side of the vertical axis. Also charted on this graph are total conflicts, represented in red. The Alternatives are as follows from left to right: existing field operations (43,969), actuated-coordinated signal control (46,988), Kadence SPF disabled (41,223), and Kadence SPF enabled (43,864). The total conflicts are charted as follows: existing field operations (1104), actuated-coordinated signal control (1219), Kadence SPF disabled (1121), and Kadence SPF enabled (1179). Figure 85. Illustration. Bell Road arterial network- comparison of total travel time. This bar graph illustrates the performance measures derived on a network-wide basis during the SIL simulation analysis. The total number of conflicts is derived from SSAM. The “Existing Field Operations” represents the optimized TOD signal plan as implemented by the operating agency. Four alternatives are charted along the horizontal axis, delay time (seconds) is measured on the left side of the vertical axis and total conflicts are measured on the right side of the vertical axis. Also charted on this graph are total conflicts, represented in red. The Alternatives are as follows from left to right: existing field operations (1,917), actuated-coordinated signal control (1,966), Kadence SPF disabled (1,877), and Kadence SPF enabled (1,922). The total conflicts are charted as follows: existing field operations (1104), actuated-coordinated signal control (1219), Kadence SPF disabled (1121), and Kadence SPF enabled (1179). Figure 86. Illustration. DC Grid network. This illustration shows a 3 by 3 grid in downtown DC, comprising of nine signalized intersections. The grid network consists of 12th Street NW, 13th Street NW and 14th Street NW (which run in a North-South direction) intersecting with E Street NW, F Street NW and G Street NW (which run in an East-West direction).

508 Compliance Report

A - 15  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

Figure 87. Illustration. DC Grid network- VISSIM model. This is a screenshot of the VISSIM network of 12th Street NW, 13th Street NW and 14th Street NW (which run in a North-South direction) intersecting with E Street NW, F Street NW and G Street NW (which run in an East-West direction). Figure 88. Illustration. DC Grid network turning movement volumes. This graphic illustrates the turning volumes of the typical weekday peak hour traffic volumes traversing the grid network for 12th Street NW, 13th Street NW and 14th Street NW (which run in a North-South direction) intersecting with E Street NW, F Street NW and G Street NW (which run in an East-West direction). Figure 90. Illustration. DC Grid network- Kadence system parameters. This is a screenshot of the system parameters. Main categories include: cycle time parameters, miscellaneous, phase sequence parameters, protected permitted parameters, offset parameters, and split parameters. Figure 91. Illustration. 14th Street, E Street and Pennsylvania Avenue NW: Kadence offset adjustments. In the bar graph, simulation time is represented across the horizontal axis and controller offset (seconds) is represented on the vertical axis. The offset is charted in green. Figure 92. Illustration. 14th St, E St and Pennsylvania Ave NW- Kadence split adjustments. This is a graphical representation of the trade-offs among phases for ring 2 during the split adjustment process. Simulation time is represented across the horizontal axis and phase time (seconds) is represented on the vertical axis. Phases 5 through 8 are represented by different colors, all on the same graph. Figure 93. Illustration. Grid network: Average delay comparison. In the graph, the performance measures derived on a network-wide basis during the SIL simulation analysis. The data is juxtaposed with the total number of conflicts (obtained from SSAM) for each alternative on a secondary axis. Three alternatives are charted along the horizontal axis, delay time (hours) is measured on the left side of the vertical axis and total conflicts are measured on the right side of the vertical axis. Also charted on this graph are total conflicts, represented in red and weighted conflicts per 1000 vehicles. The Alternatives are as follows from left to right: actuated-coordinated signal control (39), Kadence SPF disabled (53), and Kadence SPF enabled (48). The total conflicts are charted as follows: actuated-coordinated signal control (1009), Kadence SPF disabled (1126), and Kadence SPF enabled (1134). ). The weighted conflicts per 1000 vehicles are charted as follows: actuated-coordinated signal control (11.1), Kadence SPF disabled (12.4), and Kadence SPF enabled (12.5). Figure 94. Illustration. Grid network- Total delay time comparison. In the graph, the performance measures derived on a network-wide basis during the SIL simulation analysis. The data is juxtaposed with the total number of conflicts (obtained from SSAM) for each alternative on a secondary axis. Three alternatives are charted along the horizontal axis, total delay (hours) is measured on the left side of the vertical axis and

508 Compliance Report

A - 16  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

total conflicts are measured on the right side of the vertical axis. Also charted on this graph are total conflicts, represented in red. The Alternatives are as follows from left to right: actuated-coordinated signal control (338), Kadence SPF disabled (463), and Kadence SPF enabled (419). The total conflicts are charted as follows: actuatedcoordinated signal control (1009), Kadence SPF disabled (1126), and Kadence SPF enabled (1134). Figure 95. Illustration. Grid network- Comparison of numbers of stops. In the graph, the performance measures derived on a network-wide basis during the SIL simulation analysis. The data is juxtaposed with the total number of conflicts (obtained from SSAM) for each alternative on a secondary axis. Three alternatives are charted along the horizontal axis, number of stops is measured on the left side of the vertical axis and total conflicts are measured on the right side of the vertical axis. Also charted on this graph are total conflicts, represented in red. The Alternatives are as follows from left to right: actuated-coordinated signal control (35,810), Kadence SPF disabled (40,789), and Kadence SPF enabled (40,046). The total conflicts are charted as follows: actuatedcoordinated signal control (1009), Kadence SPF disabled (1126), and Kadence SPF enabled (1134). Figure 96. Illustration. Grid Network- Comparison of total travel time. In the graph, the performance measures derived on a network-wide basis during the SIL simulation analysis. The data is juxtaposed with the total number of conflicts (obtained from SSAM) for each alternative on a secondary axis. Three alternatives are charted along the horizontal axis, total travel time (hours) is measured on the left side of the vertical axis and total conflicts are measured on the right side of the vertical axis. Also charted on this graph are total conflicts, represented in red. The Alternatives are as follows from left to right: actuated-coordinated signal control (797), Kadence SPF disabled (918), and Kadence SPF enabled (875). The total conflicts are charted as follows: actuatedcoordinated signal control (1009), Kadence SPF disabled (1126), and Kadence SPF enabled (1134).

508 Compliance Report

A - 17  

Enhancing Safety and Capacity in an Adaptive Signal Control System- Phase 2

508 Compliance Report

A - 18  

 

7055 Samuel Morse Dr., Suite 100 Columbia, MD 21046 www.sabra-wang.com