A Context-Aware Localized Weighting Ensemble Model for Reservoir Inflow Forecasting

Accurate reservoir inflow forecasting is essential for sustainable watershed management and low-carbon hydropower operation. Traditional fixed-weight ensemble models lack adaptability under non-stationary hydrological conditions, limiting their reliability in reservoir operation. This study proposes a Context-Aware Localized Weighting Ensemble (CALWE) framework for reservoir inflow forecasting. The framework constructs a predictive response space from heterogeneous model outputs, enabling context identification based on similarities in model prediction behaviors. A localized weighting strategy is then employed to adaptively determine model contributions across contexts. The framework was evaluated using daily reservoir inflow data from Xiaowan Hydropower Station in the Lancang River Basin and monthly inflow data from Xiluodu Hydropower Station in the Jinsha River Basin. Results demonstrate that CALWE outperforms individual models and conventional ensemble approaches in both cases. Compared with the best-performing individual benchmark model for each basin (i.e., SVR for Xiaowan and XGBoost for Xiluodu), CALWE achieved a relative RMSE reduction of 4.93% and an absolute NSE improvement of 0.007 for daily inflow forecasting at Xiaowan, while achieving a relative RMSE reduction of 5.32% and an absolute NSE improvement of 0.027 for monthly inflow forecasting at Xiluodu. SHAP analysis revealed scale-dependent feature contributions, with daily forecasts dominated by antecedent inflow information and monthly forecasts influenced by meteorological, land surface, and hydrological factors. These findings demonstrate that CALWE captures context-dependent inflow responses while providing interpretable insights into model predictions, thereby supporting sustainable watershed management and reservoir operation.