Abstract: In view of the disadvantage of the traditional static feedforward neural network′s poor dynamic performance， the GA-Elman multivariate annual runoff prediction model which is based on the GA optimization Elman neural network connection weights is developed in this study. The model is used in the runoff forecasting of Yamadu hydrometric station in the Ili River in Xinjiang， and the traditional Elman， BP and GA-BP traditional multivariate annual runoff prediction model are used as the contrast models. Their prediction results are compared with the results of the IEA-BP network model in one reference. Comparison of the results shows that： ① GA-Elman model′s fitting and forecasting effect is slightly better than that of the IEA-BP model， and GA-Elman model for multivariate annual runoff prediction is reasonable and feasible， which has better prediction accuracy and generalization ability； and ②with the same network structure and transfer function， the prediction accuracy and generalization ability of GA-Elman model are better than those of the traditional GA-BP model， and Elman model is superior to the traditional BP model. With the ability to adapt to the time-varying characteristics， Elman feedback dynamic recursive neural network prediction performance is better than BP network. GA can effectively optimize Elman neural network weights， so that the network prediction accuracy and generalization ability have been greatly improved.