Improving Forecasting Accuracy of Annual Runoff Time Series Using RBFN Based on EEMD Decomposition

Abstract

Annual runoff forecasting is one of the most important applications for effective reservoir management. Given the time-varying and non-linear characteristics of river runoff data, a novel hybrid model is proposed to improve the forecasting accuracy. First, the original data of runoff is decomposed into a number of intrinsic mode functions (IMFs) and one residual term using the ensemble empirical mode decomposition (EEMD) method. Then, these sub-series are modeled respectively by radial basis function network (RBFN) model. Finally, the prediction results of the modeled IMFs and residual series are summed to formulate an ensemble forecast for the original annual runoff time series. The proposed hybrid model is examined by predicting the annual runoff of Maduwang station in Bahe River, China. The comparison results indicate that the proposed hybrid model can effectively enhance RBFN approach for annual runoff series forecasting accuracy and it is superior to the commonly used model like auto-regressive integrated moving average (ARIMA) and back-propagation network (BPNN).