Reference evapotranspiration forecasting in the North Etna aquifer: a comparative analysis of statistical, deep learning, and machine learning models

Accurate forecasting of reference evapotranspiration (ET0) is essential for sustainable water resource management, particularly in drought-prone regions. This study evaluates the performance of statistical, machine learning (ML), and deep learning (DL) models in forecasting ET0 in the North Etna aquifer (Sicily, Italy), a semi-arid Mediterranean area characterized by warm, semi-arid to sub-humid conditions at mid elevations. Two ET0 estimation methods were used: Hargreaves-Samani (HS), calculated monthly from long-term meteorological data collected at four weather stations, and FAO Penman-Monteith (FAO-PM), obtained as monthly records from the SIAS regional agrometeorological database. The forecasting models applied include statistical approaches (Linear Regression, Exponential Smoothing, Prophet), machine learning methods (Support Vector Regression, Random Forest, Extreme Gradient Boosting), and deep learning architectures (Long Short-Term Memory, Gated Recurrent Unit, and Temporal Convolutional Network). Model performance was evaluated using standard error metrics: RMSE, MAPE, R2, and NSE. Results indicate that statistical models perform well for HS-based ET0, which relies on a limited set of climatic variables. In contrast, ML and DL models outperform FAO-PM ET0, which incorporates more complex meteorological inputs such as wind speed and humidity. Among DL models, Long Short-Term Memory and Temporal Convolutional Network show superior ability to capture long-term temporal dependencies, making them suitable for extended sequence forecasting. This comparative analysis highlights the importance of aligning model complexity with data requirements and ET0 formulation. Findings offer practical insights for improving irrigation scheduling and water planning in Mediterranean environments facing increasing climate variability.