River flow forecasting using an integrated approach of wavelet analysis and artificial neural networks

Abstract

The need for accurate river flow forecasting model has grown rapidly in the past decades for achieving better risk-based water resources planning due to issues like water demand increase or climate change. In this paper a hybrid Wavelet-Neural Networks (WNN) is developed to predict daily river flow. WNN is one of the most reliable recent methods for hydrological time series predictions. 30 years of daily stream flow and rainfall data from Dingo road station on Harvey River, Western Australia are used in this study. Both rainfall and runoff time series are decomposed into multi-frequency time series by using the Harr and Daubechies wavelet No5 (db5), then the wavelet coefficients are imposed as input data to feed-forward back propagation ANN. The best structure of ANN is chosen by trial and error to reach best daily stream flow forecasting. Comparing the results with those of the single ANN model indicates that the performances of WNN are more effective than ANN in terms of selected performance criteria.