Abstract: Amiing at the lack of consideration of using the physical cause approach to forecasting long-term runoff,a BP(Back P ropagation) neural newtork model is developed for autumn season runoff forecasts.The model considers those physical factors affecting the runoff process.The factors include the preceding atmospheric circulation,ocean and historical runoff.Using the principal component analysis,the autumn runoff can be quantitatively related to a set of physical predictors including amtospheric circu lation,sea surface temperature,and antecedent precipitation and runoff.The model is applied to the Dan jiangkou Reservoir.Observed September and October runoff during the period 1956-2000 are used for the model calibration,while observation of 2001-2008 are used for the model validation.The results show that the stability of the model utilizing the physical predictors is generally favorable.On average,a 30% miprovement can be achieved in September and October runoff forecasting comparing to the method of using only antecedent precipitation especially during the drought and flooding years.