There are many methods for medium-and long-term runoff forecasting, such as the traditional methods of time series, multiple linear regression and etc.. Although this kind of methods are easy to use, they often have deviation in forecasting precision if the forecasting objects supply fewer samples or the factors are unreasonably chosen. This paper introduces an improved Elman neural network, which has been used in the runoff forecasting for optimized regulation of the Fengtang Reservoir. The improved Elman network has been compared with regression procedure and BP network. The results show that the method can not only raise calculating efficiency, but also increase forecasting precision.