Watershed hydrological simulation is a measure to simulate the real world hydrological processes based on hydrological models. A large number of conceptual hydrological models have been invented driven by different generalizations of the identical hydrological process. The advantages of different models can be investigated by comparison of different hydrological models. Ensemble forecasting is a means that integrates several model results to obtain more reliable and stable predictions. Arithmetic Mean, Weighted Averaging and Bayesian Model Averaging are three common ensemble forecasting methods. In this study, seven hydrological models were applied and compared in three semi-humid watersheds in the northern China to explore the applicability of different models in each watershed. Bayesian Model Averaging scheme were used to integrate these models to compare the advantages of different combinations, and the application effect of Bayesian Model Averaging scheme was studied. The results show that models with saturation-excess mechanisms have a good simulation effect in semi-humid watersheds. Traditional models are improved based on the characteristics of different watersheds, and the accuracy of simulation is higher after improvement. The certainty forecasting results and probabilistic forecasting results are provided by Bayesian Model Averaging scheme. The integrated result of few good-simulated models shows poor stability. Bayesian Model Averaging scheme can take advantages of each model efficiently and provide high-precision forecasting results by increasing the number of integrated models properly.