To evaluate the application effectiveness of the mass-conserving long short-term memory(LSTM) and support vector regression (SVR) models. Furthermore, the SHAP (Shapley additive explanations) method was applied to explore the contribution of meteorological features to the runoff prediction of the MC LSTM model and to analyze the interpretability of the model. The results show that, without the input of historical runoff observation data, the simulation accuracy of the MC LSTM and LSTM models is superior to that of the SVR model, and the MC LSTM model slightly outperforms the LSTM model during the flood season. When historical runoff observation data are included, the three models show high accuracy in short term simulations. As the lead time increases, the accuracy of the MC LSTM model declines gradually, but the model remains stable and performs better than the LSTM and SVR models in medium to long term runoff simulations. The main influencing factors of runoff prediction in the Liyuan and Wunonglong basins are temperature and precipitation, respectively. These differences in dominant factors reflect the distinct characteristics of runoff generation processes in different basins, and the interpretability analysis of the MC LSTM model helps clarify the mechanisms of runoff generation.