Advanced Machine Learning Approaches for Accurate Prediction of Reservoir Evaporation in a Tropical River Basin

Abstract

Accurate estimation of reservoir evaporation is critical for water resources management in tropical regions where evaporative losses are substantial. This study compares the following machine learning models: Least Squares Support Vector Regression (LS-SVR), Random Forest (RF), and XGBoost for monthly evaporation prediction at Hirakud Reservoir using 33 years of hydro-meteorological and operational data (1985–2017) for training and 7 years (2018–2024) for independent validation, with inputs including temperature, precipitation, humidity, wind speed, solar radiation, inflow, and release. Feature engineering incorporated nonlinear transformations and moving averages to represent temporal dependencies. Random Forest achieved the best performance (NSE 0.90 training, 0.76 testing; RMSE 8.27 MCM), outperforming XGBoost (NSE 0.69, RMSE 9.36 MCM) and LS-SVR (NSE 0.66, RMSE 9.80 MCM), and reliably captured seasonal and inter-annual variability without structural drift. Pre-monsoon months (March–May) exhibited the highest evaporation, averaging 72.65 MCM per month, 64 percent higher than winter losses. Solar radiation emerged as the dominant driver (40.1 percent relative importance), with reservoir releases significantly influencing evaporation during peak demand periods. The results demonstrate the robustness of ensemble tree-based models, particularly Random Forest, for operational reservoir evaporation forecasting and climate-adaptation planning in tropical reservoir systems.