Abstract: To improve the prediction effect of traditional runoff prediction model for stochastic time series, a forecast model of runoff based on wavelet decomposition and Arima error correction is proposed to achieve higher predictionprecision in this paper. The wavelet decomposition method is employed to decompose and reconstruct runoff time series, and smooth the non-stationary and random runoff time series. After data pre-processing, the runoff forecast model is built based on relevance vector machine (RVM), the improved particle swarm optimization (IPSO) algorithm is used for optimization, and finally the fitting errors are corrected by Arima model. Case study demonstrates that the average predictive errors of SVM model, RVM model and the proposed model are 8.60%, 9.02%, and 3.64%, respectively. Results prove that wavelet decomposition and reconstruction of time series could effectively enhance prediction precision; meanwhile, Arima error correction also has sound effect. The proposed model is of higher precision with the standard SVM model and RVM model, and therefore is feasible in engineering practice.

Key words: runoff prediction, wavelet decomposition, relevance vector machine, prediction accuracy, Arima error correction