Enhancing daily runoff forecasting in hydropower basins with a voting ensemble model using historical data

Enhancing high-precision daily runoff forecasting in hydropower basins is crucial for effective water management and flood early warning systems (FEWS). Additionally, evaluating the potential of machine learning (ML) models still needs improvement. Hence, this study initially assesses five individual ML models – multilayer perceptron (MLP), support vector regressor (SVR), random forest (RF), extreme gradient boosting (XGB), and catboost regressor (CBR) – to predict daily runoff. Subsequently, a voting ensemble (VE) model was developed to enhance the accuracy of the single models. The VE model improves performance by reducing root mean squared error (RMSE) values by 1–6% (year-round), 2–14% (dry season), and up to 7% (flood season) compared to the individual ML models. SHAP (SHapley Additive exPlanations) analysis was also applied to interpret model predictions, highlighting the most influential features. Overall, the VE model significantly improves forecast precision, making it especially effective in hydropower-affected basins, and its predictions support flood modelling and FEWS in downstream areas.