An explainable deep learning approach for hydropower production forecasting: Evidence from 116 plants in China

Accurate hydropower production forecasting is beneficial for improving water use efficiency and power system flexibility. However, the conventional two-stage method (reservoir inflow forecast, followed by generation scheduling) performs unsatisfactorily in data-scarce river basins, as it fails to capture complex rainfall–inflow–power relationships without extensive data. Additionally, purely data-driven models often lack generalizability and interpretability, limiting their application in large-scale systems. To this end, we propose a hybrid Bi-LSTM-MHSA-EKFC framework that integrates Bi-LSTM, multi-head self-attention, and an enhanced Kalman filter correction for monthly hydropower production forecasting. Using SHAP based interpretability analysis, we identify key factors influencing hydropower production forecast. Finally, we statistically examine how forecasting accuracy correlates with installed capacity and regulation capability. China's 116 hydropower plants were selected as case studies. Results show that: (1) the proposed framework achieves high predictive accuracy across regions, especially in Southwest China (median NSE = 0.83); (2) hydropower generation is most sensitive to the previous 1–2 months of power output and inflow, and to 3–6 months ahead air temperature and precipitation; (3) Large hydropower plants (>300 MW), designed primarily for power generation or equipped with large regulating storage capacity, tend to have better predictability compared with small/run-of-river plants. Therefore, this framework provides a valuable tool for hydropower generation scheduling.