Artificial Neural Networks Based Prediction for Thermal Comfort in an ...

3 downloads 0 Views 4MB Size Report
Abstract-- A predictive models were developed to determine the thermal comfort level for the academic classroom by using artificial neural networks (ANNs).
International Conference and Utility Exhibition 2014 on Green Energy for Sustainable Development (ICUE 2014) Jomtien Palm Beach Hotel and Resort, Pattaya City, Thailand, 19-21 March 2014

Artificial Neural Networks Based Prediction for Thermal Comfort in an Academic Classroom T. Songuppakarn1, W. Wongsuwan1, and W. San-um1.

Abstract-- A predictive models were developed to determine the thermal comfort level for the academic classroom by using artificial neural networks (ANNs). The paper reports experimental and theoretical analysis on a problem of achieving a desired thermal comfort level. The proposed method focused on the classical artificial (feed forward) neural networks (ANN) and the time-series NARX feedback neural networks to achieve the thermal comfort assessed using the predicted mean vote (PMV). The field measurements were conducted in a selected classroom of the Thai-Nichi Institute of Technology (TNI), Thailand. The predicted PMV agreed well with tested PMV data. Therefore, the results would be further demonstrating the feasibility and performance of the approach to achieve the classroom thermal comfort. Index Terms-- ANN, NARX, PMV, Thermal Comfort, Buildings Energy Management.

I. NOMENCLATURE bj ck dl

f cl

The bias for neurons zj The bias for the neurons yk The bias of the neuron in the output layer The ratio of body surface area covered by clothes to the naked surface

area fy ,fz The activation functions hc The convective heat transfer coefficient Hai

I cl

M Pa

The relative humidity RH (%) The clothing thermal resistance (m2◦C/W) The metabolic rate (W/m2) The partial water vapor pressure (Pa)

PMVm

PMV calculated from Fanger's model

PMVq PMV got from questionnaires

Tai The air temperature (oC) Tcl The surface temperature of clothing (oC) Tr

Va W wkj wji wjk yk zj

Mean radiant temperature (oC) The air velocity (m/s)

External work (W/m2) The weight from neuron yk The weight from neuron xi to neuron zj The weight connecting a neuron from the second hidden layer to the output layer The output of the neural network model The hidden layer with J neurons

This work was supported in part by the research grant from the research and academic service department, Thai-Nichi Institute of Technology. 1 Master Program in Engineering Technology (MET), Graduate School, and Faculty of Engineering, Thai-Nichi Institute of Technology (TNI), Bangkok, Thailand, 10250.

II. INTRODUCTION In Thailand, primary energy consumption in household and commercial building represents about 40% of the total energy consumption [1]. The academic building consumed annual electricity per usage area approximately 30 – 170 kWh/m2.yr (averaged 37 kWh/m2.yr), where contribution of airconditioned area approximately 27%. Typically, the airconditioning system of academic (schools, universities, colleges) buildings consumed electricity about 63 – 165 kWh/yr per square meter or air-conditioned area, and averaged to 76.2 kWh/m2.yr. The contribution of energy usage for airconditioning was 53%. There were more than 160 universities/colleges/institutes of undergraduate and graduate levels in Thailand, which most of large portion of academic buildings were constructed before 2000. Therefore, there is room of energy savings potential using appropriated building energy management measures. Thermal comfort is largely a state of mind that the level of comfort is often characterized using the ASHRAE thermal sensation scale. The average thermal sensation response using the ASHRAE thermal sensation scale is called the predicted mean vote (PMV), which is the focused predictive output in this research paper. To achieve thermal comfort maintenance and energy consumption minimization in most conditions are conflicted so that optimization method are required to find appropriate solutions over time. In comparison, ANN was considerably competitive to the other multi – agent control systems; Fuzzy logic, Neuro – fuzzy systems [3]. An intelligent computation in buildings by the management of indoor environment including user preferences; Fuzzy systems, synergistic neuro – fuzzy techniques, Adaptive fuzzy PD and fuzzy PID and ANN controllers, illustrated the comfort conditions. ANN is applicable to a better control of the building’s heating system, in order to control the indoor temperature of a solar building [4]. The average energy consumption of a precision inverter air conditioning (A/C) system in comparison with a conventional A/C system was investigated by T. Sookchaiya et al.[5], which they found that the system could provide people in Thailand the human thermal comfort and health. Therefore, the authors aim to develop an adaptive control algorithm of the

International Conference and Utility Exhibition 2014 on Green Energy for Sustainable Development (ICUE 2014) Jomtien Palm Beach Hotel and Resort, Pattaya City, Thailand, 19-21 March 2014

conventional A/C system based on the prediction of the thermal comfort level using ANN. III. OVERVIEW OF ARTIFICIAL NEURAL NETWORKS The neural networks concept was inspired by the basic functioning mechanisms of human brain, gathering information through a learning process. The original artificial neural networks (ANN) structure is depicted in Fig.1, comprising of two parts; an input layer and general neurons in the hidden and output layers. The output a is a function of weights W , input p and bias b . Input

General Neuron

p1 p2 p3 p4 p5

pk

W1,1

n . . .

f

a

b W1,R

a= f(Wp + b) Fig. 1. Original artificial neural network structure [6]

In this paper, the adaptive learning ability of the neural networks was applied to solve the complex problem in estimating PMV values of the air-conditioned space in an academic building. The author would like to develop few simple measurements of the indoor air conditions; i.e temperature, relative humidity and air velocity, so that the PMV values of the occupants in the classroom can be easily estimated. If the PMV lines in the acceptable range referred to the ASHRAE thermal sensation scale, it would be not necessary to change the room air conditions. In contrary, temperature and air flow rate adjustment might be required. The paper focused on two types of ANN models; the typical feed-forward neural network architecture using backpropagation learning technique, and the nonlinear autoregressive model with exogenous inputs (NARX) model. The NARX-type model is a recurrent dynamic network commonly used in time-series nonlinear modeling of the dynamic system [7]. The thermal comfort of the occupants in the air-conditioned room was considerably dynamic system. The predictive mean vote (PMV) of human in the room could vary upon the transient room air conditions and inside activities. Therefore, NARX model was used for modeling the PMV behavior of the classroom. The PMV output depends not only on the PMV input to the network but also on the current and/or previous inputs, outputs, or stages of the network.

IV. DESCRIPTION OF EXPERIMENTAL SET UP The experiments were conducted in a classroom rooms in a C-building, at the faculty of engineering, the Thai-Nichi Institute of Technology (TNI), Thailand. The classroom size is 7 x 8.5 x 3 m3, divided into 7 x 7 grids as shown in Fig.2, led to 49 measuring locations. The measurement was taken at the center of each small area which is defined by grid points. The wireless temperature and RH sensors were placed on the work plane (or classroom table) or about 0.7 m from the floor level. The air velocity probe was located nearby the sensors. Measurements were taken at least 0.5 m away from walls and columns and grid points were positioned with equal spacing (Fig.2). This standard classroom of TNI was designed for maximum occupants about 80 students. During two experiments (1st and 2nd experiment), there were 42 and 60 students sitting in the classroom, for example the case of 42 students is depicted in Fig.2. All the seated persons (students) would face the front wall. The right side wall was an external wall having single glass windows installed along this envelope. There were two doors at the left side wall (next to the corridor), 1 double door and 1 single door, which were normally closed. There were three hanged Fan Coil Unit (FCU) installed closed to the right side wall. The split type air-conditioning systems were turned on during 3 hours of the class. The field measurements of the indoor air conditions including temperature, relative humidity and air velocity, were conducted for 3 hours. For the two experiments in the same classroom, measured data of air temperature and relative humidity using wireless sensor were recorded every 2 minutes from 9:30 – 12:30 hr and 13:00 – 15:00 hr, respectively. The temperature sensor device, developed by IES laboratory at TNI, was designed based on Arduino microcontroller. The temperature and humidity sensor model DHT11 was exploited. Such a DHT11 sensor is cost-effective and provides a calibrated digital signal output with 8-bit microcontroller using with power supply of 3.5 V. The measuring temperature range is 0 oC to 50 oC with accuracy of ±0.5 oC while the measuring relative humidity range is 20% to 90% with accuracy of ±1%. Fig. 3 depicts the implementation of the temperature sensor device in which the temperature sensor senses the room temperature and sends the data to the microcontroller prior to data collection using MATLAB through personal computer interface. The primary set values for the room temperature is 25oC. The wireless T vs. RH sensors was probably calibrated with standard Thermo-Anemometer probe is AN100 model from EXTECH instruments, and the psychrometer. The air velocity was measure at all grid cells by using the probe. The measured indoor air conditions data was used in calculating PMVs at all cells using Fanger’s equation (1)[8]. Moreover, questionnaires were filled by all occupants to evaluate their thermal, humid and air movement sensation, which consequently determine PMVs. The thermal, humid and air movement sensation scale suggested by ASHRAE is

International Conference and Utility Exhibition 2014 on Green Energy for Sustainable Development (ICUE 2014) Jomtien Palm Beach Hotel and Resort, Pattaya City, Thailand, 19-21 March 2014

depicted in Table 1. The sensation vote of thermal ranges from +3 to -3 (hot to cold), of humid from +2 to -2 (humid to dry), and of air movement from +2 to -1 (too breezy to too still). 8.5 m

Instructor desk

FCU-1

FCU-2

FCU-3

(1,7)

(2,7)

(3,7)

(4,7)

(5,7)

(6,7)

(7,7)

(1,6)

(2,6)

(3,6)

(4,6)

(5,6)

(6,6)

(7,6)

(1,5)

(2,5)

(3,5)

(4,5)

(5,5)

(6,5)

(7,5)

(1,4)

(2,4)

(3,4)

(4,4)

(5,4)

(6,4)

(7,4)

(1,3)

(2,3)

(3,3)

(4,3)

(5,3)

(6,3)

(7,3)

(1,2)

(2,2)

(3,2)

(4,2)

(5,2)

(6,2)

(7,2)

(1,1)

(2,1)

(3,1)

(4,1)

(5,1)

(6,1)

(7,1)

Tai, Va Hai

LEFT SIDE WALL

hc is the convective heat transfer coefficient, as a BACK WALL

FRONT WALL

Tcl is the surface temperature of clothing, as a function of Tcl = f ( M ,W , Pa, I cl , f cl , Tcl , Tr , hc , Tai ) as given by (2)

7m

Double Door

Tai (oC) and Tr (oC) are the air temperature and mean radiant temperature.

EXTERIOR RIGHT SIDE WALL

Glass windows

Pa (Pascal) is the partial water vapor pressure

Single Door

Fig. 2. Experimental room dimension and locations of instrument

function of hc = f (hc* ,Va ) given by (3) and (4). Coefficients

a 0 to a11 are, 0.303, 0.036, 0.028, 3.05e-3,

5733, 6.99, 0.42, 58.15, 1.7e-5, 5867, 0.0014, 34, 3.96e-8, respectively. Tcl = 35.7 − 0.028( M − W ) − I cl [(3.96e − 8) f cl [(Tcl + 273) 4 (2) − (Tr + 273) 4 ] + f cl hc (Tcl − Tai )],

and

hc* hc =  12.1 Va

if hc* > 12.1 Va

(3)

* c

if h < 12.1 Va

where (4)

hc* = 2.38(Tcl − Tai )1 / 4 Fig. 3. Wireless temperature and relative humidity sensors TABLE 1 THERMAL, HUMID, AND AIR MOVEMENT SENSATION

Thermal Sensation hot warm slightly warm neutral slightly cool cool cold

Value +3 +2 +1 0 -1 -2 -3

Humid Sensation humid slightly humid just right slightly dry dry

[9]

Air Movement Sensation too breezy breezy just right too still

PMV index by Fanger [8] depending on 6 factors: clothing insulation, air temperature and humidity, air velocity, the metabolic rate, and the mean radiant temperature, was used in predicting the average vote of group of occupants on the thermal sensation scale. PMV is a function of six variables that can be measured or estimated, PMV = f ( M ,W , Pa, f cl , Tcl , Tr ) given by (1),

(

PMV = a0e

− a1 M

)

+ a2 [( M − W ) − a3 {a4 − a5 ( M − W ) − Pa}}

− a6 {( M − W ) − a7 } − a8 M ( a9 − Pa ) − a10 ( a11 − Tai ) (1) 4

4

− a11 f cl {(Tcl + 273) − (Tr + 273) } − f cl hc (Tcl − Tai )],

where

M (W/m ) and external work, 2

W (W/m ) are the metabolic rate and 2

Va (m/s) is the air velocity I cl (m2◦C/W) is the clothing thermal resistance f cl is the ratio of body surface area covered by clothes to the naked surface area, is defined by (5)

1.00 + 1.290 I cl f cl =  1.05 + 0.645I cl

if I cl ≤ 0.078

(5)

if I cl > 0.078

However, the mean radiant temperature,

Tr is a quantity

which is hard to measure, therefore, it is assumed to equal the measure indoor temperature.

The flow chart in Fig.4 shows iterative procedure to develop the PMV values based on Fanger’s model. Equation (2) to (4) was solved iteratively until a prescribed degree of convergence was attained or a maximum number of iterations were reached.

Pa (Pascal), the vapor pressure of moisture is easily function of the relative humidity of the air ( H ai ) by means of Antoine’s equation (6).

Pa = 10(H ai e(16.6536 - 4030.183/ (Tai + 235))

(6)

Derivations of all equations mentioned above, resulted in the PMV as a function of 4 variables, by assuming similar metabolic rate, and external work for the student in the

International Conference and Utility Exhibition 2014 on Green Energy for Sustainable Development (ICUE 2014) Jomtien Palm Beach Hotel and Resort, Pattaya City, Thailand, 19-21 March 2014

classroom. Therefore, the paper simply determines PMV as a function of only 4 variables; indoor air temperature, relative humidity, velocity, and the clothing thermal resistance as in (7).

Research Methodology

(7)

PMV = f (Tai , H ai , Va , I cl )

Field Measurements

Start

Tai, Hai, Va

Input: Tai, Hai, Va, M, Icl, Tr Hai,Tai Yes

Calculate PMV by Fanger’s equation

Icl Icl 12.1(Va)(-½)

Calculate Pa, Eq.(6)

hc=hc* fcl

Pa

fcl

Calculate PMV, Eq.(1)

No

hc=12.1(Va)

Tcl,new > Tcl,i

(Tcl,new-Tcl,i) < 0.2(Tcl,i)

PMVq

NARX modeling

NARX modeling

PMVNARX

PMVANN (-½)

Comparison

PMVNARX Comparison

ANN modeling and prediction

Solve for Tcl,new , Eq.(3)

No

Yes

Eq.(3)

Calculate PMV from Questionnaires PMVm

Typical ANN modeling

fcl=1.05+0.645(Icl)

Yes

Questionnaires

Discussion and Analysis Yes

No

Tcl

Calculate hc, Eq.(2) hc

End

Fig. 4. Flow chart of PMV determination by Fanger’s Equation

V. METHOD The overall research methodology illustrated as flow chart in Fig. 5 comprises of three parts, (1) experiments by field measurements, (2) questionnaires, and (3) development of neural networks models (typical ANN and NARX models), including prediction and comparison of the resulted PMV. The field measurements were conducted in an actual classroom during its normal operation, to gather indoor air conditions data in determining of Fanger’s PMVm. However, the questionnaires were also developed to survey thermal/humid/air movement sensation vote from the seated persons in the classroom to identify the PMVq accordingly. There are two approaches of ANN modeling and prediction; typical (or classical) ANN model and NARX model. Each has different input-output data patterns; the typical ANN model needs 3 inputs (Tai, Hai, Va) and 1 output (PMVm) to predict PMVANN. Otherwise, the NARX model needs only 1 input (PMVm or PMVq) to predict PMVNARX in the time-series accordingly. The predicted PMV from different models are then compared and discussed. The author expect that, in the next research phase, the optimum ANN model would be used in the adaptive control algorithm to maintain thermal comfort and to achieve energy saving.

Fig. 5. Overall research methodology

A. Field Measurements The field measurements were carried out as follows; (1) Design of experiments in measuring of the airconditioning classroom to collect sufficient data for determining the state-of-the-art PMVm from Eq.(1). (2) Design, construct, preliminary testing, calibration, adjustment and preparation of an appropriate temperature and humidity sensors. (3) Design the questionnaires to survey information on temperature, humid and air movement sensation from the seated persons in the classroom. (4) Conduct 2 experiments by field measurements in the same classroom (C-402) for 2 days consecutively to collect experimental data. The 1st experiment (in Jan 2014) lasted from 13:00 – 15:00 hr, with 42 seated persons. The 2nd experiment was conducted in the next day with 60 seated persons during 9:30 – 12:30 hr. All three A/C systems were turned on in both experiments. (5) Data selection and screening statistically to prepare sufficient data patterns for training/validation/testing of typical ANN and NARX model. (6) Compile and analyze surveyed data from questionnaires to verify the calculated PMVm. B. ANN Modeling In the theoretical part, there are two types of ANN modeling approach, (1) classical or typical artificial neural networks (ANN) modeling and (2) NARX modeling.

International Conference and Utility Exhibition 2014 on Green Energy for Sustainable Development (ICUE 2014) Jomtien Palm Beach Hotel and Resort, Pattaya City, Thailand, 19-21 March 2014

In case of classical ANN model, each pattern contained 3 input variables (Tai, Hai, Va) and 1 output (PMVm). The inputoutput of structure of the model was specified, and then the two design parameters must be determined; the number of neurons and the training patterns. For the number of neurons, a number of trial was carried out in the range [2, 7] as suggested by related research work [10-11]. Several training patterns were trial until the simple and appropriated one, the TRAINLM training function with Levenberg-Marquardt back propagation type was satisfied.

better accurate. When the series-parallel NARX network (open loop) training process is successfully done, the feedback loop is closed as shown in Fig. 6. During the open loop training procedure, the Levenberg-Marquardt algorithm was used in weight coefficients adjustment. The best network configuration of NARX model had 1 x 10 x 1 topology, where there was one input layer, one hidden layer with ten nodes, and one output layer with one node, respectively.

Tai

In the classical ANN model, the k outputs (y1,…..yk,…..,yK) were transformed from the I inputs (x1,…..,xi,…..xI) through the hidden layer with J neurons (z1,…..,zj,…..,zJ). The output of the neural network model, yk can be determined in (8);

Hai

PMV

Va 3 nodes 4 nodes

(8) Input layer

(9) Where wkj is the weight from neuron yk, ck is the bias for the neurons yk, wji is the weight from neuron xi to neuron zj, bj is the bias for neurons zj, wjk is the weight connecting a neuron from the second hidden layer to the output layer, and dl is the bias of the neuron in the output layer. fy and fz are the activation functions, which are normally nonlinear functions. Among nonlinear functions, the Sigmoid shape activation functions are normally used and defined as (10) (10) The ANN should result in the smallest discrepancy between the Fanger’s and calculated PMVs, so that various configurations for the ANN were used. Number of neurons in the input and output layers are fixed by number of independent variables and the target PMV as 3 and 1, respectively. By trial and error, the number of neurons in the two hidden layer was finally 7 and 4. In ANN modeling, the data must be divided into 2 groups; the first one was for training and the rest was for testing of the model based on balancing of statistical data. The training an testing data set should have approximately similar minimum to maximum ranges and average PMV values as in the main data set. In case of the NARX model, the Fanger’s PMV values were employed as inputs of the model for the prediction of the next time-series ahead of PMV. Firstly, the efficient training approach using a seriesparallel architecture of the NARX network was conducted. Therefore, the real PMV output can be used instead of feeding back the estimated output, so that the input to the network is

7 nodes 2 Hidden layers

Output layer

Fig. 6. Resulted Artificial Neural Networks (ANN) architecture 10 nodes

W

PMV(t)

W

+

b

Input layer Hidden layer

W

b

+

PMV(t+dt)

Output layer

Fig. 7. Time series NARX feedback neural networks architecture

VI. RESULT AND DISCUSSION The measured data from two tests in the experimental classroom, the surface plots of air temperature, relative humidity, and air velocity are depicted in Fig. 8 and 10, referring to the room 7 x 7 grids. The surface plots shows that the indoor air temperatures are in the range of relative humidity 50% to 80%, and air velocity about 0.03 to 2.04 m/s (1st experiment) and 0.03 to 1.2 m/s (2nd experiment), respectively (Fig. 8(a) and 10(a)). Considering influence of relative humidity, the occupants were mostly just right in the range of relative humidity 60% to 80% at air velocity in the

International Conference and Utility Exhibition 2014 on Green Energy for Sustainable Development (ICUE 2014)

24

25-26

23

24-25

22

23-24 21-22

80-100

40

60-80

20

40-60

0

20-40

6 7

7

6

2

1

Axis Title

22-23

7

60

5

4 5

3

6

5

4

1

Room grid

4

3

2

5

6

7

0-20

2.5 2 1.5

2-2.5

1

1.5-2 1-1.5

0.5

1

0.5-1

3

2

0

4

4

7

6

5

1

Room grid

2

3

5

6

7

0-0.5

5 6

-3

7

1

-4

2

3

4

-1-0 -2--1 -3--2 -4--3

-5

3 5

-1

2

1

0 -0.5

4

0-1

6

5

7

-1.5

1

7

2 3 4

-2

6

PMV from questionnaires

1 1

PMV (Fanger's model)

(c ) Fig. 8. Measured room air conditions (a) Temperature, (b) Relative humidity, and (c) Air velocity, from the 1st experiment

0

2

3

4

5

6

7 -0.5-0 -1--0.5

-2

-1.5--1

-2.5

-2--1.5

-3

-2.5--2 -3--2.5

-5--4

Room grid

40-50

30 20

30-40

10

20-30

0

10-20

3 4 6

5

1

Room grid

(a)

2

4

3

6

5

7

0-10

(b)

Room air velocity (m/s)

1.2 1 0.8

1-1.2

0.6

0.8-1 0.6-0.8

0.4 0.2 0

0.4-0.6

7 1

0.2-0.4

5 2

3

4

0-0.2

3 5

Room grid

6

7

1

-1.5

-1--0.5

-2

-1.5--1 -2--1.5

-2.5

-2.5--2

-3

-3--2.5

-3.5

1 0.5 0

0.5-1

-0.5 -1

1

7

-0.5-0

5

7

6

6

2

5

3

4

4

3

1

2

PMV from questionnaires

1

7

4

5

-1

6

2

PMV (Fanger's Model)

0 -0.5

3

(c) Fig. 10. Measured room air conditions (a) Temperature, (b) Relative humidity, and (c) Air velocity, from the 2nd experiment

-1.5

2

3

4

5

6

7

0-0.5 -0.5-0 -1--0.5 -1.5--1

-2

-2--1.5

-3.5--3

(a)

(b)

Fig. 11. PMV distribution within the experimental classroom (a) calculated by Fanger’s Equation, (b) based on Questionnaires, from the 2nd experiment

(b) Room air velocity (m/s)

(a)

-1

T

50-60

40

7

7

6

5

4

3 Room grid

2

60-70

50

Room grid

80

4

3

2

21

4

3

6

5

24-24.4

7

70-80

60

2

2

24

1

24.4

100

1

26-27

2

25

24.4-24.8

70

Room grid

3

Room air relative humidity (%)

26

1

Room Air Temperature (oC)

27

24.8-25.2

24.8

room air relative humidity (%)

25.2

The PMVs values calculated by Fanger’s equation (1) and those analyzed from the occupants questionnaires are illustrated and compared in Fig. 9 and 11. The seated students felt mostly “cool” and “slightly cool” during the 1st experiment (Fig. 9). In the 2nd experiment, PMVm values implied that most of the occupants felt “cool” and “slightly cool” (Fig. 11), however, the PMVq inferred mostly neural. Nonetheless temperature, relative humidity and air velocity had significant effect to thermal sensation vote. Large discrepancies between PMVq and PMVm might be from state of mind in favoring cooled condition of Thai student, which thermal acceptability for the classroom at percent of vote “0” are 42%. The results from questionnaires shows that all the occupants activities in the classroom were seating, writing and reading, and wearing trousers, long – sleeved shirt, long – sleeved sweater and T – shirt, resulted in the metabolic equivalent of task about “Met 1.0”.

25.2-25.6

80

1

25.6

1

range of 0.03 to 0.5 m/s and temperature in the ranges of 24.0 °C to 24.9 °C (Fig. 8(b) and 10(b)). The humid acceptability of the occupants increased with relative humidity, however, the value lower than 80% for this case. The occupants felt just right in the ranges of air velocity about 0.03 to 0.5 m/s, getting votes as “breezy” and “too breezy” at the higher air velocity in the ranges of 1.0 to 2.5 m/s (Fig. 8(c) and 10(c)). The highest air movement acceptability can be obtained in the ranges 0f 24.0 °C to 24.9 °C.

Room air temperature (oC)

Jomtien Palm Beach Hotel and Resort, Pattaya City, Thailand, 19-21 March 2014

Room grid

(a) (b) Fig. 9. PMV distribution within the experimental classroom (a) calculated by Fanger’s Equation, (b) based on Questionnaires, from the 1st experiment

During typical ANN training, 49 data patterns were used that each training data set contained 3 measured data; indoor air temperature, relative humidity and velocity, and one calculated PMVm using Fanger’s Equation (1), and about 10 patterns were used for model testing. In the same way, the NARX training, validation and testing data sets had similar number of data patterns. Training and testing data was selected by considering similar range and average to the whole set of data. In comparison of 62 data sets, between the PMVs calculated by Eq. (1) and the PMVs from the questionnaires, there was about one-half (35 data sets) having acceptable discrepancy ( PMVq − PMVm ≤ ±1 point), implying the effect of individual thermal comfort sensation on the PMV. Also, among these verified set, there was 15 sets having thermal/humid/air movement as PMV value about -1 (slightly cool, slightly dry and too still) and 0 (neural and just right). Table 2 summarizes statistical data of the training data required by ANN modeling; temperature ( 24.9 ± 0.4 oC ), relative humidity ( 70.3 ± 3.6% ), air velocity ( 0.24 ± 0.28 m / s ) and PMVm ( − 1.58 ± 0.56 ).

International Conference and Utility Exhibition 2014 on Green Energy for Sustainable Development (ICUE 2014) Jomtien Palm Beach Hotel and Resort, Pattaya City, Thailand, 19-21 March 2014

TABLE 2 STATISTICS OF THE TRAINING DATA PATTERNS FOR ANN

Input/Output

Unit o

Indoor air temperature Relative humidity Air velocity PMV (Fanger's model)

C % m/s (-)

Min

Max

Avg.

Stdev

24.0 61.0 0.03 -3.48

25.5 75.0 1.20 -0.76

24.9 70.3 0.24 -1.58

0.4 3.6 0.28 0.56

To validate the network, the regression plot must be created showing the relationship between the outputs of the network and the targets. If the training were perfect, the network outputs and the targets would be exactly equal. Once the training/validation/testing processes for ANN modeling were completed and satisfied the desired conditions, all weights and biases would be reported, and those obtained are summarized in Table 3-5, for the case of classical ANN PMV model, and the time-series NARX PMV model. These weights and biases shall be used in predicting the PMV of the classroom. The ANN model required measured parameters; air temperature, relative humidity and velocity, to predict the PMV. However, the NARX model required historical PMV to forecast the next time-series PMV values. For the ANN model testing, the R2 resulted from the feedforward neural network type TRAINLM by the tested PMV data was in acceptable level about 0.9736. However, for the NARX model testing of PMV, the R2 obtained are 1 in both cases of PMVq and PMVm. Therefore, the predicted values of the PMV (predicted by NARX) agreed well with the PMV values from Fanger’s model and from the questionnaires. TABLE 3 WEIGHT AND BIAS OF EACH NEURON IN FEED FORWARD CLASSICAL ANN PMV MODEL j Wj,i=1 (IW1,1)

1

2

3

4

5

6

7

-1.6863

-2.4890

-3.0759

-1.5992

-0.9022

-2.2372

-0.1188

Wj,i=2 (IW1,1)

-2.6737

-0.9216

-2.3589

-1.8775

2.8173

1.5970

-2.1083

Wj,i=3 (IW1,1)

-2.5711

1.8402

-1.1133

-0.9746

-3.0137

-2.1925

-1.6313

Wk=1,j (IW2,1)

-0.7485

-0.8181

0.0450

-0.7636

-0.7085

0.5019

0.6080

Wk=2,j (IW2,1)

0.6567

2.4553

-1.8155

-0.6052

1.2878

-0.8344

1.2723

Wk=3,j (IW2,1)

-1.3937

0.8790

-0.7772

2.2171

1.5879

-1.8535

1.7898

Wk=4,j (IW2,1)

-0.1295

-1.2613

-3.0653

0.1125

-3.1163

0.5914

-1.4479

Wl=1,k (IW3,2)

0.6372

2.4571

-2.3409

-0.7002 0.36559

2.2753

-0.50821

bj (Bias to layer 1) 0.29826 Ck (Bias to layer 2)

1.6776

dl (Bias to layer3)

1.0201

l

1

k

1

-1.4543

-1.6344

1.8883

-0.12009

-0.79864

0.062323

2

3

4

TABLE 4 WEIGHT AND BIAS OF EACH NEURON IN TIME-SERIES NARX PMV (FANGER EQUATION) MODEL 1

2

3

4

5

6

7

8

9

Wj,i=1

0.2106

-1.6203

-1.1068

-0.3686

-0.7553

0.6010

-0.0275

1.0304

-2.1287

0.5961

IWinput feedback

j

-1.5335

0.2318

1.3782

-0.8007

-0.9507

-0.6107

3.1320

1.2546

0.4144

-0.3593

10

Wk=1,j (LW2,1)

0.0006

0.0009

-0.0046

-0.0869

-0.7232

1.3372

-0.0065

0.3925

-0.1319

-0.0002

LW1,3 (LayerWeightFeedBack)

-0.2167

-0.7635

-2.0606

0.6756

0.9627

0.6050

-2.6492

-1.2336

-0.3020

-0.3057

(LW2,2)

2.6514

2.2571

1.6074

0.2256

-0.0202

0.0106

0.2721

-0.0081

0.4619

-2.1765

Bias input

2.0470

1.7351

1.9674

0.0201

-0.2902

-0.8775

1.9566

1.6206

-3.2152

3.1862

BiasFeedback

0.2556

TABLE 5 WEIGHT AND BIAS OF EACH NEURON IN TIME-SERIES NARX PMV (QUESTIONNAIRES) MODEL 1

2

3

4

5

6

7

8

9

10

Wj,i=1

j

0.6559

-0.9463

2.1320

1.5378

-0.4221

0.8429

-1.5324

0.4221

-1.9314

0.0212

IWinput feedback

2.2049

0.1974

-0.8015

0.6319

-0.8857

0.1098

-2.1996

-1.8837

0.2718

0.7090

Wk=1,j (LW2,1)

0.0000

0.0000

0.3012

0.0000

-0.0373

1.2254

0.0000

-0.0373

0.0000

-0.0381

LW1,3 (LayerWeightFeedBack)

-4.5791

-1.8355

0.8015

0.3120

2.3398

-0.1098

2.1637

0.4296

1.6283

-0.7268

(LW2,2)

-1.8986

-0.6669

0.0000

-1.2061

-2.2569

0.0000

-2.5333

2.2569

-1.8231

-0.0263

Bias input

-0.7782

2.6203

-1.4901

1.9591

-0.2280

0.4266

0.1451

0.2280

-1.2424

-4.6892

BiasFeedback

-0.2548

VII. CONCLUSION Thermal comfort in an air-conditioned classroom is a major concern, especially in case of academic buildings of the universities and/or institutes in large city. The thermal comfort should be maintained while energy consumption in airconditioning system must be taken into account. The PMV of the occupants in the classroom should be included into the control system of the air-conditioning system. The artificial neural networks modeling approach was exploited in the research for PMV prediction as compared to PMV values obtained from questionnaires. Two type of ANN models were trial; a classical ANN model, and NARX model. The optimum topology of ANN and NARX models are 1 x 7 x 4 x 1, and 1 x 10 x 1, respectively. The weights and biases for further model prediction are reported in the paper. The predicted PMV results from ANN and NARX models agreed well with either Fanger’s PMVs or questionnaire PMVs. In the next research phase, it expected to develop an adaptive control algorithm of the conventional A/C system based on the predicted thermal comfort level called PMV using ANN approach. VIII. ACKNOWLEDGMENT The authors gratefully acknowledge the contributions of Mr. Somsak Thaicharoen, the MET master student, for his assistance in instrument and data measurement. The authors also wish to acknowledge the support of Mr. Jirayu Peetakul his support to develop the temperature sensors and ANN modeling source code in MATLAB. IX. REFERENCES [1] EPPO (2011). Thailand 20-year energy efficiency development plan (2011 – 2030), Ministry of Energy. [2] EPPO (2000). National Energy Policy Office, Thailand Energy Strategy and Policy, May 2000, [Online]. Available: http://www.eppo.go.th/encon/Strategy/encon-EnergyStrategy-E.html [3] A.I. Dounis*, C. Caraiscos .(2009). Advanced control systems engineering for energy and comfort management in a building environment – A review. Renewable and Sustainable Energy Reviews, 13, 1246 – 1261. [4] A.A. Argirioua,*, I. Bellas – Velidisb, C.A. Balarasa. (2000). Development of a neural network heating controller for solar buildings. Neural Networks, 13, 811 – 820. [5] Thammanoon Sookchaiya, Veerapol Monyakul, Sirichai Thepa*. (2010). Assessment of the thermal environment effects on human comfort and health for the development of novel air conditioning system in tropical regions. Energy and Buildings, 42, 1692 – 1702. [6] Artificial Neural Network. [Online]. Available: http://www.gotoknow.org/posts/472128 [7] S. Fischer, P. Frey, and H. Druck. (2012). A comparison between state-ofthe-art and neural network modeling of solar collectors. Solar Energy, 86, 3268-3277.

International Conference and Utility Exhibition 2014 on Green Energy for Sustainable Development (ICUE 2014) Jomtien Palm Beach Hotel and Resort, Pattaya City, Thailand, 19-21 March 2014

[8] P.M. Ferreirab,c,*, A.E. Ruanoa,c, S.Silvaa, E.Z.E. Conceicaoa. (2012). Neural networks based predictive control for thermal comfort and energy savings in public buildings. Energy and Buildings, 55, 238 – 251. [9] American Society of Heating, Refrigerating and Air Conditioning Engineers (ASHRAE), “Thermal Environmental Conditions for Human Occupancy,” ANSI/ASRAE Standard 55 – 1992, ASHRAE, Atlanta, 1992. [10] Meysam Karamirada, Mahmoud Omida,*, Reza Alimardania, Hossein Mousazadeha, Seyyed Navid Heidarib . (2013). ANN based simulation and experimental verification of analytical four – and five – parameters models of PV modules. Simulation Modelling Practice and Theory, 34, 86 – 98.1751-1757. [11] T. Kazanasmaz, M.Gunaydin, S. Binol. (2009). Artificial neural networks to predict daylight illuminance in office building. Building and Environment, 44, 1751-1757.

X. BIOGRAPHIES Tharntip Songuppakarn graduated from Kasetsart University and holds a Bachelor’s degree from the faculty of Engineering, in which Chemical Engineering major, Bangkok, Thailand. She has successfully completed the 2010 Japan Overseas Development Corporation Human Resources Program – Getting to Know Japan Commissioned by the Ministry of Economy, Trade and Industry (METI) of Japan. She has studied in Master of Engineering Technology Program at Thai– Nichi Institute of Technology (TNI). Her employment experience includes the Trans Thai – Malaysia (Thailand) Ltd., Bureau Veritas (Thailand) Ltd., Energy Scope Co.,Ltd, Thailand Institute of Nuclear Technology (TINT) and Inoac Automotive (Thailand) Co., Ltd. Her special fields of interest included the alternative energy of technologies, on the modeling and simulation of the PV system for building application or other related renewable energy or thermal system for building facilities. Dr. Wipawadee Wongsuwan was born in Bangkok in 1973. She received the B.Eng. degree in chemical engineering from Kasetsart University, Thailand in 1994, and the D.Eng. degree for the dissertation on solar assisted solid adsorption and chemical heat pump systems for energy upgrading, from the Asian Institute of Technology (AIT), Thailand, in 2003. From 1995 – 2010, she was a lecture at the mechanical engineering department, faculty of engineering, Chiang Mai University (CMU), Chiang Mai, Thailand, where she was an Assistant Professor. She was involved research projects on PV assisted roof solar collector, reduction of heat transfer through building envelop by double wall, heat transfer enhancement in thermal equipment, etc. In 2010, she joined the faculty of engineering, Thai-Nichi Institute of Technology (TNI), has started the Master Program in Engineering Technology (MET), and has responsible as a director of the MET graduate program of the faculty engineering. Dr. Wimol San-Um was born in Nan Province, Thailand in 1981. He received B.Eng. Degree in Electrical Engineering and M.Sc. Degree in Telecommunications in 2003 and 2006, respectively, from Sirindhorn International Institute of Technology, Thammasat University, Thailand. In 2007, he was a research student at University of Applied Science Ravensburg-Weingarten, Germany. He received PhD in LSI Designs in 2010 from the Department of Electronic and Photonic System Engineering, Kochi University of Technology, Japan. He is a lecturer at Computer Engineering Program, Faculty of Engineering, Thai-Nichi Institute of Technology. His areas of research interests are analog integrated circuit designs, involving chaotic oscillators and switched-current circuits, and on-chip testing design, involving DFT and BIST techniques.