Abstract
Air temperature plays a critical role in estimating agricultural water requirements, hydrological processes, and the climate change impacts. This study aims to identify the most accurate forecasting model for daily minimum (T(min)) and maximum (T(max)) temperatures in a semi-arid environment. Five machine learning models-linear regression (LR), additive regression (AR), support vector machine (SVM), random subspace (RSS), and M5 pruned (M5P)-were compared for T(max) and T(min) forecasting in Gharbia Governorate, Egypt, using data from 1979 to 2014. The dataset was divided into 75% for training and 25% for testing. Model input combinations were selected based on best subset regression analysis, result shows the best combination was T(min(t-1)), T(min(t-3)), T(min(t-4)), T(min(t-5)), T(min(t-6)), T(min(t-7)), T(min(t-8)) and T(max (t-1)), T(max (t-2)), T(max (t-3)), T(max (t-4)), T(max (t-5)), T(max (t-6)), T(max (t-8)) for daily minimum maximum air temperature forecasting, respectively. The M5P model outperformed the other models in predicting both T(max) and T(min). For T(min), the M5P model achieved the lowest root mean square error (RMSE) of 2.4881 °C, mean absolute error (MAE) of 1.9515, and relative absolute error (RAE) of 40.4887, alongside the highest Nash-Sutcliffe efficiency (NSE) of 0.8048 and Pearson correlation coefficient (PCC) of 0.8971. In T(max) forecasting, M5P showed a lower RMSE of 2.7696 °C, MAE of 1.9867, RAE of 29.5440, and higher NSE of 0.8720 and R² of 0.8720. These results suggest that M5P is a robust and precise model for temperature forecasting, significantly outperforming LR, AR, RSS, and SVM models. The findings provide valuable insights for improving decision-making in areas such as water resource management, irrigation systems, and agricultural productivity, offering a reliable tool for enhancing operational efficiency and sustainability in semi-arid regions. The Friedman ANOVA and Dunn's test confirm significant differences among temperature forecasting models. Additive Regression overestimates, while Linear Regression and SVM align closely with actual values. Random Subspace and M5P exhibit high variability, with SVM differing significantly. For maximum temperature, Random Subspace and M5P perform similarly, while SVM remains distinct.