Reservoir inflow forecasting using Voting Ensemble model: A case study at A Luoi hydropower, central Vietnam

Accurate reservoir inflow forecasting is critical for real-time water management in monsoon-dominated basins. This study develops a weighted Voting Ensemble model to predict daily inflow to the A Luoi hydropower reservoir in central Vietnam using multi-station rainfall and lagged inflow data. Five machine learning models MLP, RF, KNN, XGB, and Ridge Regression, were trained on a unified feature set containing current and lagged rainfall inputs and three runoff memory terms, and subsequently combined using performance-based weights derived from time-series cross-validation errors. Evaluation using MSE, RMSE, and NSE shows that the ensemble outperforms all standalone learners, reducing RMSE by 12–25% and improving NSE from 0.70–0.91 (best individual models) to 0.92 on the test set. SHAP analysis is also used to explain model predictions and highlight the most influential features. During an independent verification period, the ensemble maintained strong performance (NSE ≈ 0.98), accurately capturing rising and recession limbs and minimizing peak-flow underestimation. These results demonstrate the robustness and operational feasibility of weighted ensemble learning for short-term inflow forecasting, offering valuable support for reservoir operation, flood mitigation, and water allocation in data-rich reservoir systems.