Combined Wavelet-Neural Network Model for Intermittent Stream Flow Prediction

Abstract

Achieving accurate intermittent river flow forecasting, plays a key role in water resources and environmental management. Water demands are increasing while surface water availability is likely to decrease in Western Australia. Understandably, reliable information on current and future water availability is essential for properly manage and share the limited water resources. Forecasting intermittent stream flow is quite limited due to the complexity of fitting models to their time series as they do not have flow for some intervals. In this paper Wavelet-Neural Networks (WNN) technique is studied to reach accurate and reliable daily river flow prediction. WNN is based on combination of wavelet analysis and Artificial Neural Network (ANN), which is one of the most reliable recent hybrid methods for non-stationary hydrological time series predictions. Daily stream flow and precipitation historical data from Northam weir station on Avon River, Western Australia are used in this study. The observed stream flow and rainfall time series are both decomposed by Daubechies4 and Coiflet1 Wavelet transforms. Then the sub-series are added up to develop new time series for imposing as input data to the multilayer perceptron neural networks (MLP). Comparing the results of different wavelet neural networks with those of the single ANNs model indicates that preprocessing data with discrete wavelet transform have significantly improved artificial neural in terms of selected performance criteria.