Comparison of Wavelet-Based ANN and Regression Models for Reservoir Inflow Forecasting

Abstract

The present study demonstrates the capability of two preprocessing techniques such as wavelets and moving average (MA) methods in combination with feed-forward neural networks—namely, back propagation (BP) and radial basis (RB) and multiple linear regression (MLR) models—in the prediction of the daily inflow values of the Malaprabha reservoir in Belgaum, India. Daily data on 11 years of rainfall, inflow, and streamflow at an upstream gauging station have been used. The observed inputs are decomposed into subseries using discrete wavelet transform with different mother wavelet functions, and then the appropriate subseries is used as input to the neural networks for forecasting reservoir inflow. Model parameters are calibrated using 7 years of data, and the remaining data are used for model validation. More statistical indices have been used to determine the optimal models. Optimum architectures of the wavelet neural network (WNN) models are selected according to the obtained evaluation criteria in terms of the Nash–Sutcliffe efficiency coefficient, root mean squared error, and correlation coefficient. The result of this study has been compared by developing two standard neural network models and a multiple linear regression (MLR) model and MA. The results of this study indicate that the WNN model performs better compared to artificial neural network (ANN) and MLR models in forecasting the inflow hydrograph effectively. The study only used reservoir inflow data from one area, and further studies using data from various areas may be required to strengthen these conclusions.