University of Tabriz Journal of Hydrogeomorphology 2383-3254 4 11 2017 08 23 The Prediction of the Flood Peak Discharge Using a Wavelet Neural Network The Prediction of the Flood Peak Discharge Using a Wavelet Neural Network 149 168 6719 FA Hamidreza Babaali - Assistant Professor of Civil Engineering, Islamic Azad University of Khorramabad Reza Dehghani PhD candidate of Water Structure, Faculty of Agric., University of Lorestan, Khorramabad, Iran, (Corresponding author), Journal Article 2016 12 13 Introduction Flood is one of the hazardous natural disasters that causes loss of life and financial problems every year. Therefore, scientists have tried to assess the quantitative variability of this phenomenon as much as possible. In this study, the recorded data in Kahman Aleshtar watershed area, which is located in Lorestan province, was used to investigate the precision of the different flood peak discharge prediction models. In addition, the wavelet and artificial neural network models were selected for the modeling of the flood peak discharge and the results were compared to examine the accuracy of the studied models. Methodlogy Daily flood peak discharges of the basin in Kahman station, which were applied for the calibration and validation of the models, were selected and observed. For this purpose, maximum daily precipitation rate, at a daily scale and between the years 2001-2012, and flood peak discharge were respectively used as the input and output parameters. The wavelet-based neural network which was based on the combination of the wavelet theory and neural networks were created. Indeed, it has the benefits and features of the neural networks and the charm, flexibility, strong mathematical foundations, and the analysis of the multi-scale wavelets. The combination of the wavelet theory with the neural network concepts for the creation of the wavelet neural network and feed-forward neural shock can be a good alternative for estimating the approximate nonlinear functions. Feed-forward neural network with sigmoid activation function is in the hidden layer. While at the nerve shocked wavelet, the wavelet functions as the activation of the hidden layer feed-forward networks are considered, in both these networks and scale wavelet, the transformation parameters are optimized with their weight. Artificial neural networks inspired by the brain's information processing systems, designed and emerged into. To help the learning process and with the use of the processors called neurons, there was an attempt to understand the inherent relationships among the data mapping, the input space, and the optimal space. The hidden layer or layers, the information received from the input layer, and the output layer are the processing and disposal. Based on the artificial neural network structure, its major features are high processing speed and the ability to learn the pattern, the ability to extend the model after learning, the flexibility against unwanted errors, and no disruption to the error on the part of the connection due to the weight distribution network. The first practical application of the synthetic networks with the introduction of the multilayer perceptron networks was consultation. For training this network, back propagation algorithm is used. The basis of this algorithm is based on the error correction of the learning rule. That consists of two main routes. By adjusting the parameters in the MLP model, error signal and input signal occurs. Determining the number of the layers and neurons is the most important issue in simulation with the artificial neural network. The criteria of the correlation coefficient, the root mean square error, and the mean absolute error were used to evaluate and compare the performance of the models. Results The results showed that both models in a structure, consisting of 1 to 4 delay, gives better results than any other structure. In addition, based on the results of the evaluation criterion, the model which was used to wavelet neural network model, was the most accurate (R=0.921), and the lowest root mean square error RMSE=0.005m3/s and the lowest average absolute error MAE=0.003m3/s the validation phase is capable. Conclusions Wavelet neural network model outperformed the artificial neural network. Consequently, it can be effective in forecasting the daily flood peak discharge. It can also facilitate the development and the implementation of the surface water management strategies. Finally, predicting the piver flow process is a major step in water engineering studies and water resources' management. Introduction Flood is one of the hazardous natural disasters that causes loss of life and financial problems every year. Therefore, scientists have tried to assess the quantitative variability of this phenomenon as much as possible. In this study, the recorded data in Kahman Aleshtar watershed area, which is located in Lorestan province, was used to investigate the precision of the different flood peak discharge prediction models. In addition, the wavelet and artificial neural network models were selected for the modeling of the flood peak discharge and the results were compared to examine the accuracy of the studied models. Methodlogy Daily flood peak discharges of the basin in Kahman station, which were applied for the calibration and validation of the models, were selected and observed. For this purpose, maximum daily precipitation rate, at a daily scale and between the years 2001-2012, and flood peak discharge were respectively used as the input and output parameters. The wavelet-based neural network which was based on the combination of the wavelet theory and neural networks were created. Indeed, it has the benefits and features of the neural networks and the charm, flexibility, strong mathematical foundations, and the analysis of the multi-scale wavelets. The combination of the wavelet theory with the neural network concepts for the creation of the wavelet neural network and feed-forward neural shock can be a good alternative for estimating the approximate nonlinear functions. Feed-forward neural network with sigmoid activation function is in the hidden layer. While at the nerve shocked wavelet, the wavelet functions as the activation of the hidden layer feed-forward networks are considered, in both these networks and scale wavelet, the transformation parameters are optimized with their weight. Artificial neural networks inspired by the brain's information processing systems, designed and emerged into. To help the learning process and with the use of the processors called neurons, there was an attempt to understand the inherent relationships among the data mapping, the input space, and the optimal space. The hidden layer or layers, the information received from the input layer, and the output layer are the processing and disposal. Based on the artificial neural network structure, its major features are high processing speed and the ability to learn the pattern, the ability to extend the model after learning, the flexibility against unwanted errors, and no disruption to the error on the part of the connection due to the weight distribution network. The first practical application of the synthetic networks with the introduction of the multilayer perceptron networks was consultation. For training this network, back propagation algorithm is used. The basis of this algorithm is based on the error correction of the learning rule. That consists of two main routes. By adjusting the parameters in the MLP model, error signal and input signal occurs. Determining the number of the layers and neurons is the most important issue in simulation with the artificial neural network. The criteria of the correlation coefficient, the root mean square error, and the mean absolute error were used to evaluate and compare the performance of the models. Results The results showed that both models in a structure, consisting of 1 to 4 delay, gives better results than any other structure. In addition, based on the results of the evaluation criterion, the model which was used to wavelet neural network model, was the most accurate (R=0.921), and the lowest root mean square error RMSE=0.005m3/s and the lowest average absolute error MAE=0.003m3/s the validation phase is capable. Conclusions Wavelet neural network model outperformed the artificial neural network. Consequently, it can be effective in forecasting the daily flood peak discharge. It can also facilitate the development and the implementation of the surface water management strategies. Finally, predicting the piver flow process is a major step in water engineering studies and water resources' management.

https://hyd.tabrizu.ac.ir/article_6719_7acae4f7ea5ee6b81525e06d9f03dcb9.pdf