Performance of artificial intelligence algorithms for reference evapotranspiration forecasting across different climatic zones

Abstract

Accurate forecasting of reference evapotranspiration (ET₀) is important for irrigation planning and water resources management, particularly in regions characterized by contrasting climatic conditions. Although various artificial intelligence models have been applied to ET₀ prediction, systematic comparisons of advanced hybrid architectures across distinct climatic settings remain limited. This study evaluated one-month-ahead forecasting of monthly ET₀ at three synoptic stations representing extremely arid, semi-arid, and arid coastal conditions in Iran. Seven forecasting approaches were compared: Autoregressive Integrated Moving Average-Long Short-Term Memory (ARIMA-LSTM), Support Vector Machine implemented for regression (SVM), Support Vector Machine optimized using a Genetic Algorithm (SVM-GA), Support Vector Machine-Random Forest (SVM-RF), Transformer, Temporal Convolutional Network (TCN), and a hybrid TCN-Transformer model. Model performance was evaluated using the coefficient of determination (R²) and root-mean-square error (RMSE). The TCN-Transformer model consistently produced the most accurate forecasts across the three investigated stations, with R² values ranging from 0.95 to 0.99 and dimensionless normalized RMSE values ranging from 0.06 to 0.11. The conventional SVM-based models generally produced lower forecasting accuracy, particularly at the arid coastal station. These results indicate that the ET₀ series at the arid coastal station was more difficult for the evaluated models to forecast. Overall, the findings demonstrate the potential of combining convolution-based feature extraction with attention-based modeling for accurate one-month-ahead monthly ET₀ forecasting under the climatic conditions investigated in this study.