RBF Artificial Neural Network Model based on Wavelet De-noising for Hydrological Time Series Simulation and Forecast

Hydrological system is a nonlinear unstable complex system influenced by various factors. Hydrological time series are the expressions of hydrological system via data, in whichnoiseis inevitable. According to nonlinear problems and noise pollution in hydrological system, RBF neural network based on wavelet de-noising (WD-RBF-ANN) was appliedtosimulate and forecasthydrological time series. The technology of wavelet de-noising by soft-threshold was introduced, in which the wavelet function waschosen to analyze the series and Heuristic SUREmethod was selected toeliminate error in data. The improved RBF artificial neural network was investigated to predict time series after wavelet de-noising. The structure of model was created via self-learning ability according to time series sample, while the relevant parameter values were optimized by mean square error (MSE) criterion. The simulation and forecast results of WD-RBF-ANN model were compared with RBF-ANN and ARIMA model. Illustrated by the case ofannual runoff series of Huayuankoustation and Lijin station, annual precipitation series of Beijing and Nanjing, comparative analysis showed that WD-RBF-ANN has superiority in accuracy and the precipitation results in various regions could provide reference to water resources management.