Modeling Temporal Persistence in Daily Pan Evaporation Using Interpretable Machine Learning

Accurate prediction of daily pan evaporation (Ep) is critical for irrigation management and water resource planning, yet it remains challenging due to strong temporal variability and persistence. Most available models are based primarily on contemporaneous meteorological variables and do not account for the inherent memory of evaporation. This paper systematically explores the impact of temporal memory on daily Ep by explicitly separating evaporation persistence from meteorological memory through a machine learning framework. A Categorical Boosting (CatBoost) model was used to analyze multi-year data from a semi-arid to sub-humid area in western-central India. The model used lagged Ep and meteorological variables at 1-, 3-, 7-, and 14-day timescales. The coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE) were used to evaluate model performance, while Random Forest and XGBoost were used for benchmarking. Feature importance and SHAP analyses provided interpretability. The findings indicate that the inclusion of a 1-day lag of pan evaporation substantially improves predictive performance, with RMSE decreasing to 0.83 mm. This improvement indicates strong short-term persistence in day-to-day evaporation. Longer evaporation lags degrade performance, whereas lagged meteorological variables provide modest additional improvement beyond evaporation memory. CatBoost demonstrated the most balanced performance among the evaluated models. Overall, the findings confirm that short-term evaporation persistence dominates temporal dependence in daily Ep, with meteorological memory playing a secondary role, thereby offering a physically interpretable and efficient framework for data-driven evaporation modeling.