Abstract: The aim of this research is to improve the accuracy of forecasting snowmelt runoff in the mountainous areas of western Tianshan Mountains, and to better support the development of industrial and agricultural production in the study area. The predictor, which is a key issue affecting forecast accuracy, are optimized and selected by using mutual information, correlation coefficient method, and principal component analysis method. The selected predictors are taken as input factors in RBF neural network model and combinatorial wavelet BP neural network model for comparison. Results suggest that: 1) optimized predictors selected by the mutual information method could improve forecast accuracy; 2) according to forecast results under different scenarios, the results of RBF neural network model is superior to those of combinatorial wavelet BP neural network model; 3) with relative error as the standard of accuracy evaluation, RBF neural network model with input factors selected by mutual information method could produce the optimum forecast result.

Key words: mid-term and long-term hydrological forecasting, mutual information, coefficient of correlation, principal component analysis, neural network