Abstract: To compare the prediction accuracy and interpretability of different machine?learning algorithms for reservoir outflow forecasting, this study employs observed daily inflow, outflow, and current water level data of the Three Gorges Reservoir from 2008 to 2025. Four models—Extreme Gradient Boosting (XGBoost), Back Propagation (BP) neural network, Long Short-Term Memory (LSTM) network, and Generalized Additive Model (GAM)—are selected. Their outflow prediction accuracy is compared under three hydrological scenarios: the full period, the flood season (June–October), and the dry season (November–May of the following year). The SHapley Additive exPlanations (SHAP) framework is introduced to interpret the feature contribution mechanism of each model. The results show that XGBoost achieved the highest prediction accuracy in all three scenarios（R2=0.803~0.884，RMSE=1420~3569 m3/s）,The RMSE values are all within 20% of the mean observed outflow. SHAP analysis reveals that the driving directions of XGBoost and GAM are opposite to those of BP and LSTM under high water levels, despite generally similar feature response trends across the four models. Given the misalignment between prediction performance and interpretability across scenarios, model selection must prioritize the physical plausibility of decision logic alongside accuracy metrics.