Under the analysis of the mechanism and influence elements of watershed runoff, the watershed runoff forecasting model based on the feedforward Neural Network (NN) is established in this paper. It selects seven factors as the inputs of NN, including antecedent precipitation,main runoff duration, four representative stations' rainfall intensity and area average rainfall,and runoff depth as the output.A new algorithm-hybrid GN-BFGS algorithm is introduced, which is better than the traditional BP method in calculating speed and convergence.The paper also discusses the hidden unit numbers of NN.