Integrating Seasonal Generation Patterns for Enhanced Short-Term Hydropower Output Forecasting in Power Systems

To address the challenges where traditional machine learning methods struggle to capture the nonlinear and seasonal characteristics of hydropower output, this study develops a short-term hydropower output forecasting method by integrating seasonal output characteristics. Firstly, seasonal hydropower output characteristics are extracted as key reference features by using a self-organized mapping neural network to construct a multi-dimensional input feature matrix including historical output, meteorological factors, and reference features. Then, the eXtreme Gradient Boosting (XGBoost) ensemble learning algorithm is utilized to establish a short-term forecasting model. Finally, the SHapley Additive exPlanations (SHAP) framework is introduced to quantify the marginal contributions of key features and elucidate the model's forecasting rationale from the perspective of feature attribution. Taking a provincial power grid in central China as a case study, the results indicate that the XGBoost model integrating seasonal output characteristics achieves a coefficient of determination (R²) of 0.919 over a 24-hour forecast horizon. Compared with the benchmark model that does not incorporate seasonal output characteristics, the proposed model improves the value of R² by 11.5% and reduces the value of Root Mean Square Error (RMSE) by 41.7%. SHAP analysis demonstrates that, apart from basic temporal features such as historical output and time, the contribution of seasonal output features ranks at the forefront. This suggests that these features provide a key reference for the intraday output process to the model, effectively correcting forecast deviations. This method achieves short-term hydropower output forecasting with both high accuracy and high interpretability, providing technical support for the flexible operation and safe management of high-proportion renewable energy power grids.