Flood Water Level Prediction and Tracking Using ... - IEEE Xplore

2014 IEEE 10th International Colloquium on Signal Processing & its Applications (CSPA2014), 7 - 9 Mac. 2014, Kuala Lumpur, Malaysia

Flood Water Level Modeling and Prediction Using NARX Neural Network: Case Study at Kelang River Fazlina Ahmat Ruslan#1, Abd Manan Samad*2, Zainazlan Md Zain #3, Ramli Adnan#4 #

Faculty of Electrical Engineering Dept. of Surveying Science and Geomatics, Faculty of Arc., Planning and Surveying Universiti Teknologi MARA, 40450 Shah Alam, Selangor, Malaysia

*

Corresponding Email: [email protected] Abstract— Flood disaster has becomes major threat around the world because it causes loss of lives and damages to property. Thus, reliable flood prediction is very much needed in order to reduce the effects of flood disaster. Hence, an accurate flood water level prediction is an important task to achieve. Since flood water level fluctuation is highly nonlinear, it is very difficult to predict the flood water level. Artificial Neural Network is well known technique is solving nonlinear cases and Nonlinear Auto Regressive with Exogenous Input (NARX) model is one class of Artificial Neural Network model. Thus, this paper proposes flood water level modeling and prediction using Nonlinear Auto Regressive with Exogenous Input (NARX) model to overcome the nonlinearity problem and come out with an advanced neural network model for the prediction of flood water level 10 hours in advance. The input and output parameters used in this model are based on real-time data obtained from Department of Irrigation and Drainage Malaysia. Results showed that NARX model successfully predicted the flood water level 10 hours ahead of time.

Levenberg-Marquardt algorithm to obtain the best results. ANN also has been used to predict water flow at TadamiAgano river basin in Japan for three years duration [14]. The results showed that the developed model was able to predict water flow during flood events. Markus et al [15] successfully applied Back Propagation Neural Network as one type of ANN to predict stream flows at a gauging station in Southern Colorado, USA on monthly basis. They used snow as one of the input parameters. Nonlinear Autoregressive with Exogenous Input (NARX) is one class of ANN models. The NARX model was developed based on Autoregressive with Exogenous Input (ARX) which represents linear system identification model. The NARX models have been widely applied in various fields because it can represent any nonlinear functions. For example, Saad and Mashor [16] used three different input parameters for NARX-RLS (Recursive Least Square) algorithm to predict car speed. Each input parameters for NARX1, NARX2 and NARX3 contains 3000 data sets obtained from car sensors. The results showed NARX1 with injected fuel and two output lags as input parameters was the best model compared with three NARX-RLS model. Yunan et al. [17] applied Radial Basis Function Neural Network (RBFNN) based on NARX as identification model for essential oil extraction system. The results from residual test and One Step Ahead (OSA) fitting test showed that the proposed model successfully represent the essential oil extraction system.

Keywords—Flood Water Level Prediction; Artificial Neural Network; Nonlinear Auto Regressive with Exogenous Input (NARX)

I. INTRODUCTION In recent years, Artificial Neural Network (ANN) has been widely used for prediction and forecasting in various fields. The Artificial Neural Network concept was first proposed by McCulloch and Pitts in 1943 [1]. Later, the ANN has been subject matters among the researchers around the world after the introduction of first ANN training algorithm by Rosenblatt in way back 1958 [2]. The ability of ANN to solve complex nonlinear system without the needs of any physical knowledge of the system itself also makes ANN one of the most popular black-box models [3-6].

With the advantages of ANN and NARX model mentioned above, this paper proposed flood water level modeling and prediction using NARX structure. This paper was organized in the following manner: Section II states the related theory, Section III explains the methodology, Section IV is on results and discussion and finally, Section V is on conclusions.

ANN was found to produce satisfactorily results in hydrological fields due to its advantages in nonlinear inputoutput representation [7-11]. For example, Yiming et al. [12] successfully predicted flood events in China from 1949 to 1994. For future works, the models application can also be extended to danger forecasting. Meanwhile, with the applications of ANN model, 18 months ahead groundwater level in the Messara Valley, Greece were successfully predicted [13]. The optimal parameters were used in the model development such as the fastest training algorithm namely

978-1-4799-3091-3/14/$31.00 ©2014 IEEE

II. RELATED THEORY A. Description of NARX The Nonlinear Autoregressive with Exogenous Input (NARX) model as shown in Figure 1 is the extension of the ARX model and is given by Eq.(1);

y(t ) = f ( y(t −1),...y(t − n), u(t −1),...u(t − m)) + e(t )

204

(1)


III. METHODOLOGY

where e(t) is the white noise residuals and the next value of dependent output signal, y(t) is regressed on previous values of the output signal and previous values of an independent (exogenous) input signal, u(t).

A. Data Collection The flood location analyzed in this study was Kelang river located at Petaling bridge, Kuala Lumpur as shown in Figure 2. The most significant contribution for flood location comes from three upstream rivers which were Kelang river at Sulaiman bridge, Kelang river at Tun Perak bridge and Gombak river at Jalan Parlimen. River conditions were measured in terms of water level and this real-time data can be obtained online from the Department of Irrigation and Drainage Malaysia website www.water.gov.my.

The output predictor for NARX model is given by Eq.(2) and illustrated in Figure 1 where yˆ is the prediction of the autoregressive output signal y from previous values itself and of the exogenous input signal.

yˆ (t ) = ψ ( y (t − 1),... y (t − n ), u (t − 1),...u (t − m ))

(2) B. Data Used For NARX modeling, data used was divided into two sets namely training and testing data. The model was developed using training data and the performance was then tested using testing data. The data used for model development was in meters starting from 18/11/2010 8:20:00 am till 21/11/2010 18:20:00 am in 10 minutes time interval. This data was used in the training stage because flood events happened during that time frame and thus providing good sample of data. The data used for model testing also in meters starting from 6/3/2011 7:00:00 till 9/3/2011 17:00:00 am. This data was chosen because it fulfilled the same criterion as training data. Therefore, with these optimal conditions of training and testing data, good performance result was expected at the end of this study.

yˆ

Figure 1. NARX Series-Parallel Architecture

There are two architectures that can be applied for NARX network. They are parallel (P) and series-parallel (SP) architectures. In parallel architecture, previous predicted output signals were applied as one of input signals to predict the next output signals thus creating feedback loop between input and output signal. As for series-parallel architecture, only measured outputs were used as one of input signal and thus the feedback loop was eliminated. Series-parallel architecture offer advantages of feed-forward architecture and thus static back propagation algorithm can be used as training algorithm [18, 19]. Figure 1 shows the NARX series-parallel architecture with implementation of four input signals namely u1(t), u2(t), u3(t) and u4(t) which represents water levels at three upstream rivers and difference of water levels at flood location due to rainfall. y(t) represents water levels at downstream river or so called flood location. yˆ represents predicted water levels at flood location.

Figure 2. Location of Kelang River at Petaling bridge [22]

C. Performance Indices The performance of NARX model is calculated using best fitting criterion, Akaike’s Final Prediction Error (FPE), Loss Function (V) and Root Mean Square Error (RMSE) as given in [20]. These performance measures are used because its ability to compare the preciseness between the predicted model and the true model.

205


IV. RESULTS AND DISCUSSION

NARX Model estimates versus actual value

17.5

0.6

Actual Water Level NARX Model

17 classification of water level 1) normal - 13 to 14.5 meter 2) alert - 14.5 to 15.5 meter 3) warning - 15.5 to 16 meter 3) danger - > 16 meter = flood

delay = 65 steps 65 steps x 10 min. = 650 min.

16

650 min / 60 min = 10.833 hrs

0.2 0.1

15.5

0

15

-0.1

14 0

RMSE =0.0806 m

0.3

e r r o r (m e te r )

w a t e r le v e l(m e t e r )

0.5 0.4

16.5

14.5

Error of NARX estimates versus actual value

0.7

-0.2

65 steps

-0.3 0

50

100

150

200

250 no of data

300

350

400

450

50

100

150

200

500

250 no of data

300

350

400

450

500

Figure 4. Prediction error from the NARX model

Figure 3. Prediction result using NARX model

Figure 4 shows the predicted error from the NARX model. It can be observed that the main errors contribution comes from the water peaks situation at time steps 60-80 and 250-275 where prediction was ineffective. Other regions showing good prediction with errors close to 0. The prediction performances of NARX model are given in Table I. It can be seen that the prediction performances showing good results with errors almost zeros and Best Fit of 87 %.

Figure 3 shows the prediction result of flood water level using NARX model. Four inputs of river water levels were fed to NARX model to predict flood water level at Kelang river located at Petaling bridge. The basin information was not included as one of input parameters because from literature reviews done, physiographical factors didn’t give significant effects to prediction performance [21]. From Figure 3, it can be clearly seen that NARX model successfully predicted flood water level 10.833 hours in advance. This 10.833 hours prediction time was obtained based on 65 delay steps of NARX model whereby each steps represented 10 minutes time interval. Thus, 650 minutes subsequently equal to 10.833 hours of prediction time. Delay steps were measured during water level in normal condition. Later, after 10.833 hours, water level from upstream rivers reached downstream rivers thus causing flood events to happen. As can be observed from the figure, the NARX model seems successfully predicting actual water level movement. Exception to the condition of high frequency sections or water peak occurrences. This phenomenon is typical to any tracking or prediction results in literature.

TABLE I.

PERFORMANCE INDICES RESULT

Performance Indices

NARX Model

Best Fit (%)

87.0799

Akaike’s Final Prediction Error (FPE)

0.0110 m

Loss Function (V)

0.0109 m

Root Mean Square Error (RMSE)

0.0806 m

V. CONCLUSIONS Flood modeling and prediction using NARX was successfully developed. The flood water level at downstream river or so called the flood location are successfully predicted 10.833 hours ahead of time with good prediction results. The effects of physiographical factors such as basin area, length of main stream and mean slope were neglected in this study. Therefore for future works, NARX model can be applied to other hydrological problems such as rainfall-runoff prediction and stream flow forecasting which are new to the researchers around the world.

206


ACKNOWLEDGEMENT The authors would like to thanks and acknowledge the financial support from the Faculty of Electrical Engineering and Research Management Institute (RMI) under RIF Fund 600-RMI/DANA 5/3/RIF (887/2012), Universiti Teknologi MARA, Shah Alam.

[17]

[18]

REFERENCES [1] [2] [3]

[4]

[5]

[6] [7]

[8]

[9]

[10]

[11]

[12] [13] [14]

[15] [16]

W. S. McCulloch and W. Pitts, "A logical calculus of the ideas immanent in nervous activity," The Bulletin of Mathematical Biophysics, vol. 5, pp. 115-133, 1943. F. Rosenblatt, "The perceptron: a probabilistic model for information storage and organization in the brain," Psychological review, vol. 65, p. 386, 1958. M. Behrang, E. Assareh, A. Ghanbarzadeh, and A. Noghrehabadi, "The potential of different artificial neural network (ANN)

[19]

[20]

techniques in daily global solar radiation modeling based on meteorological data," Solar Energy, vol. 84, pp. 1468-1480, 2010. N. Mohsenifar, N. Mohsenifar, and K. Mohsenifar, "Using Artificial Neural Network (ANN) for Estimating Rainfall Relationship with River Pollution," Advances in Environmental Biology, vol. 5, pp. 1202-1208, 2011. S. Taghadomi‐Saberi, M. Omid, Z. Emam‐Djomeh, and H. Ahmadi, "Evaluating the potential of artificial neural network and neuro‐fuzzy techniques for estimating antioxidant activity and anthocyanin content of sweet cherry during ripening by using image processing," Journal of the Science of Food and Agriculture, 2013. S. Kalogirou, E. Mathioulakis, and V. Belessiotis, "Artificial neural networks for the performance prediction of large solar systems," Renewable Energy, vol. 63, pp. 90-97, 2014. F. A. Ruslan, N. K. Zakaria, and R. Adnan, "Flood modelling using Artificial Neural Network," in Control and System Graduate Research Colloquium (ICSGRC), 2013 IEEE 4th, 2013, pp. 116120. R. Adnan, F. A. Ruslan, A. M. Samad, and Z. M. Zain, "Artificial neural network modelling and flood water level prediction using extended Kalman filter," in Control System, Computing and Engineering (ICCSCE), 2012 IEEE International Conference on, 2012, pp. 535-538. F. A. Ruslan, R. Adnan, A. M. Samad, and Z. Md, "Flood Prediction Modeling Using Hybrid BPN-EKF And Hybrid ENNEKF: A Comparative Study," Journal of Engineering Research and Applications (IJERA) ISSN, vol. 2248, p. 9622. A. P. Duncan, A. S. Chen, E. C. Keedwell, S. Djordjević, and D. Savić, "Urban flood prediction in real-time from weather radar and rainfall data using artificial neural networks," Proc. Weather Radar and Hydrology, Exeter, UK, 2011. F. A. Ruslan, A. M. Samad, Z. M. Zain, and R. Adnan, "Modelling flood prediction using Radial Basis Function Neural Network (RBFNN) and inverse model: A comparative study," in Control System, Computing and Engineering (ICCSCE), 2013 IEEE International Conference on, 2013, pp. 577-581. Y. Wei, W. Xu, Y. Fan, and H.-T. Tasi, "Artificial neural network based predictive method for flood disaster," Computers & Industrial Engineering, vol. 42, pp. 383-390, 2002. I. N. Daliakopoulos, P. Coulibaly, and I. K. Tsanis, "Groundwater level forecasting using artificial neural networks," Journal of Hydrology, vol. 309, pp. 229-240, 2005. T. Egawa, K. Suzuki, Y. Ichikawa, T. Iizaka, T. Matsui, and Y. Shikagawa, "A water flow forecasting for dam using neural networks and regression models," in Power and Energy Society General Meeting, 2011 IEEE, 2011, pp. 1-6. M. Markus, J. D. Salas, and H.-S. Shin, "Predicting streamflows based on neural networks," in Water Resources Engineering, 1995, pp. 1641-1646. Z. Saad and M. Mashor, "Model Structure Selection for Speed Forecasting with Nonlinear Autoregressive with an Exogenous

[21]

[22]

207

Input," in Intelligent Systems Modelling & Simulation (ISMS), 2013 4th International Conference on, 2013, pp. 290-293. I. Yunan, I. M. Yassin, S. F. S. Adnan, and M. H. F. Rahiman, "Identification of essential oil extraction system using Radial Basis Function (RBF) Neural Network," in Signal Processing and its Applications (CSPA), 2012 IEEE 8th International Colloquium on, 2012, pp. 495-499. F. A. Ruslan, A. M. Samad, Z. M. Zain, and R. Adnan, "Flood prediction using NARX neural network and EKF prediction technique: A comparative study," in System Engineering and Technology (ICSET), 2013 IEEE 3rd International Conference on, 2013, pp. 203-208. F. A. Ruslan, A. M. Samad, Z. M. Zain, and R. Adnan, "Multiple Input Single Output (MISO) ARX and NARX model of flood prediction system: A comparative study," in Control System, Computing and Engineering (ICCSCE), 2013 IEEE International Conference on, 2013, pp. 254-258. R. Adnan, F. A. Ruslan, A. M. Samad, and Z. Md Zain, "Extended Kalman Filter (EKF) prediction of flood water level," in Control and System Graduate Research Colloquium (ICSGRC), 2012 IEEE, 2012, pp. 171-174. T. Kerh and C. Lee, "Neural networks forecasting of flood discharge at an unmeasured station using river upstream information," Advances in Engineering Software, vol. 37, pp. 533543, 2006. Google Maps. (2013). [Kelang River at Petaling bridge] [StreetMap]. Retrieved from http://maps.google.com