Correlation analysis technique was used to identify main influence factors of runoff processes for prediction river section from regional precipitation, discharge or stage of river survey section, especially when the data type of river survey sections was discharge, the combined discharge obtained on combination of lag times of runoff between survey sections and prediction section was selected as influence fac tors, and the auto-correlation analysis technique was adopted to identify influence factor from preceding discharge or stage process of prediction section. And then the influence factors were used as input of networks, and the discharge or stage of prediction section was used as output of networks, with node number of hidden layer acquire by trial and error automatically, the flood forecasting model based on BP neural networks was established. Then the established model was loaded in national flood forecasting systems. The application results show that satisfactory forecasting effects are acquired when there are some representative flood processes in the training samples.