Based on the basic principles of artificial neural networks and methods, a runoff forecasting model of hidden multilayerBP neural network was built， with runoff prediction at the Yamaha station on the Ili River as the study case. Otherwise， theconventional models based on single hidden layer BP， RBF and GRNN neural networks were structured for a comparative analysis.The results showed that ① the runoff prediction model of hidden multilayer BP neural network has a strong generalization ability，prediction accuracy is high， and algorithm is stable. The model accuracy is better than the IEA-BP neural network model， whichmeans as a new method to be used in runoff prediction， it is reasonable and feasible. ② The prediction accuracy of RBF， GRNNneural network models is higher than the conventional single hidden layer BP network model because of advantages of fastconvergence， high prediction accuracy， fewer parameters to be adjusted， and easy to fall into local minimum.