Hybrid Wavelet Transform-Neural Network Approach for Short Term and Long Term Time Series Flow Forecasting

Abstract

Accurate modeling of runoff is useful in urban and environmental planning, flood and water resources management. In this research, a hybrid model has been developed for Brahmaputra River flow forecasting based on wavelet and artificial neural network (ANN) methods. In this current study, discrete wavelet transform was linked to ANN naming Wavelet Artificial Neural Network (WANN) for flow forecasting. Ten year daily flow data from January 1990 to December 1999 of Pandu and Pancharatna stations on Brahmaputra River, which carries heavy flood in monsoon season in the North-East region of India, were used in the study. The observed flow data were decomposed (up to 7 level) to multiresolution time series via discrete wavelet transform using Daubechies wavelets of order ranging from 1 (db1) to 5 (db5). Then multiresolution time series data were fed as input to ANN to get the forecasted discharge values. Daily data were used to forecast flow values for lead times 2, 3, 4, 7 and 14 day, weekly data were used to forecast flow values for lead times 1 week and 2 week, and monthly data were used to forecast flow values for lead time 1 month. The root mean square error (RMSE), determination coefficient (R2), mean absolute error (MAE), BIAS (B), and scatter index (SI) were adopted to evaluate the model‟s performance. It was found that for all lead times WANN model has given better and consistent results compared to conventional ANN model. It was mainly because of multiresolution time series used as inputs. Also it was found that, model efficiency increases with increase in wavelet order, giving best results for db5 mother wavelet for all lead times for both the stations. Also, there has been significant impact of decomposition level on WANN model efficiency as observed in the study.