Long-term prediction for autumn flood season in Danjiangkou Reservoir basin based on OSR-BP neural network

Abstract

A new method, called OSR-BP neural network, for long-term runoff prediction is put forward in this thesis. In order to eliminate input multi-collinearity and phenomenon of overfitting of the neural network, optimal subset regression (OSR) and BackPropagation(BP) neural network is coupled to an integrated, meanwhile, the training and testing error is comprehensively considered to determine the best condition of stop training. On this basis, runoff in September and October in Danjiangkou Reservoir, is simulated from 1956 to 2000, and is predicted from 2001 to 2008 by using of OSR-BP neural network. The result shows that the stability of model is favorable and accuracy is satisfactory whether simulation or prediction, especially for forecasting the characteristics of drought and flood year.