Forecasting of daily reference evapotranspiration using random tree, ANFIS and multi-regression methods

Evapotranspiration plays a critical role in water resource management and reservoir operation; however, accurately estimating it remains challenging, particularly when meteorological data are variable. Although machine learning techniques have been widely applied in evapotranspiration modeling, the systematic influence of progressively increasing combinations of meteorological inputs on the performance of hybrid and regression-based models has not been comprehensively evaluated. In this study, reference evapotranspiration (ET) was calculated using the FAO-56 Penman–Monteith method and used as the benchmark variable. Daily meteorological data, including solar radiation (SR), air temperature (T), relative humidity (RH), and wind speed (WS), were obtained from a station located at Lake Hartwell, South Carolina (USA). Four structured input combinations (SR; SR + T; SR + T+RH; SR + T+RH + WS) were developed to assess the incremental contribution of meteorological variables. The predictive performance of Multiple Linear Regression (MLR), Random Tree (RT), Adaptive Neuro-Fuzzy Inference System (ANFIS), and three polynomial regression variants (Q-MR, PQ-MR, and I-MR) was evaluated and compared with empirical Hargreaves–Samani (HS) and Turc methods using RMSE, MAE, APE, and the coefficient of determination (DC). Results demonstrated a consistent improvement in model accuracy with increasing input dimensionality. In the four-input configuration, Q-MR4 and I-MR4 achieved the highest predictive performance (RMSE = 0.073–0.074 mm/day; DC = 0.998), followed by ANFIS4 (RMSE = 0.121 mm/day; DC = 0.991). In contrast, empirical methods showed substantially higher error levels (RMSE up to 1.549 mm/day) and reduced agreement, particularly during peak evaporative conditions. Overall, the findings indicate that incorporating non-linear and interaction terms within regression frameworks provides a computationally efficient yet highly accurate alternative to more complex hybrid models for ET₀ estimation in reservoir-based climatic settings.