Synthetic streamflow generation using Artificial Neural Network

Abstract

Abstract In this study, Artificial Neural Network (ANN) models have been developed as an alternative to classical stochastic auto regressive model for monthly streamflow generation at multiple sites of a river basin. In the ANN model development a three-layer fully connected feed-forward network was developed using log-sigmoid and pure-line activation function in hidden layer and output layer respectively. Monthly neural network models have been designed for simultaneous generation of streamflows in two sites at a time. Training of those models has been done in supervised environment with back-propagation algorithm. The network has been selected based on its minimum mean square error. Available historical flow sequences of a river basin have been partitioned into three parts at 2:1:1 ratio. The first partitioned portion of the historical flow sequences were used to generate 50 sets of flow sequences with the help of multi-site first order Markov model. The generated flow data were used to train the ANN model. The trained ANN model were simulated with the second partitioned historical flow data. The simulated results have been compared with the third partitioned historical flow data in terms of statistical parameters- means, standard deviations, serial correlations and cross correlations. The ANN models so developed have been applied at four sites of Damodar River Basin: (i) two sites (Konar and Panchet) on River Damodar and (ii) two sites (Tilaiya and Maithon) on river Barakar, tributary of river Damodar, in India using available 56 years (1960–2015) of historical flow data. Statistical parameters: means, standard deviations, serial correlation coefficients and cross correlations (lag-zero, lag-one and lag-two) of historical and ANN generated series are compared. Those results indicate that all those statistical parameters as obtained from historical flow series are well preserved in the flow series generated by the ANN models.