In order to improve the precision and generalization ability of runoff forecast, the paper put forward integrated forecast model of BP neural network based on 3 kinds of basic improved algorithm. The autocorrelation analysis method and ADF unit root test method are used to determine input vector and smooth model of runoff time series. According to the standard of BP neural network algorithm with slow convergence, easy to fall into local minimum defect respectively, the paper improved the standard of BP neural network algorithm by using adaptive momentum gradient method, conjugate gradient method and Levenberg-Marquardt method, constructed 3 kinds of improved BP neural network model to predict the annual runoff at Nagara River Wenshan Donghu lake hydrological station,and contructed GA-BP model as comparison model. It integrated the results of the single forecastmodel by using weighted average method. The results show that the average absolute value of relative error of annual runoff of the Nagara River from 2001 to 2005 forecasted by integrated model is 4.67%, the maximum absolute value of relative error is 7.11%, the accuracy and generalization ability are better than that of the single model and GA-BP model. The integration overcame the disadvantages of single model prediction accuracy is not high and the error is not stable, has good prediction accuracy and generalization ability and is an effective method to improve prediction of runoff forecast.