Abstract: To improve the accuracy of runoff simulation and enhance the interpretability of hydrological models, this study constructs a coupled model that integrates SWAT and deep learning. The Bayesian method is conducted for the optimization of parameters and the Stacking ensemble strategy is used to integrate the outputs of individual model during model construction in order to reduce the uncertainties of high-flow predictions. The results show that the coupled model is better than any single model in terms of higher coefficient of determination (R²) and Nash-Sutcliffe efficiency coefficient (E_NS) and lower mean absolute error (E_MA). Specifically, the R², E_NS and E_MA obtained from SWAT-LSTM are close to 0.766, 0.761 and 4.93 m³/s, respectively, in the testing phase. By introducing the SHAP-based interpretability analysis, the results also indicate that the performance of LSTM relies more heavily on precipitation because of its characteristic of sequential memory, whereas the Transformer focuses on lateral flow through its self-attention mechanism, revealing the differentiated mechanisms by which different deep learning models capture hydrological processes. Moreover, the performance of the Stacking-based ensemble model is close to an average of 0.81 in terms of E_NS in the testing phase. This study provides a modeling approach for intelligent hydrological simulation and forecasting that maintain acceptable accuracy and interpretability.