Accurate and reliable medium and long-term forecast results of reservoirs are of great significance for guiding the optimal allocation of water resources in the intake area. In this paper, the SARIMA model, SVM model,XGBoost model and RF model were first selected to construct the monthly runoff forecasting scheme of the reservoir.Based on the physical mechanism of meteorological factors, the key predictors were screened on the basis of the cause analysis and random forest importance ranking and were input into four single models. Then, on the basis of comparative analysis of the advantages and disadvantages of each model, a combined forecasting scheme is constructed in two ways of linear and nonlinear combination. The results show that the simulation results of the RF model are the best among the four single models, and the simulation accuracy of the SARIMA model increases with the increase of the inflow runoff;the combined forecast model is better than any single model. The nonlinear combination of neural networks can effectively improve the simulation accuracy of the verification period and increase the generalization ability of the model.