A Comparative Evaluation of Regression-Based Machine Learning Models for Daily Maximum Temperature Prediction and Heatwave Assessment

Accurate prediction of daily maximum temperature is important for understanding and assessing heatwave risk, particularly in regions where extreme heat poses growing environmental and public health concerns. This study presents a comparative evaluation of five regression-based machine learning models-Linear Regression, K-Nearest Neighbours (KNN), Support Vector Regression (SVR), Random Forest, and XGBoost-for daily maximum temperature prediction and heatwave assessment. The methodology was designed around a temperature-based framework in which daily maximum temperature (Tmax) was used as the primary input variable, followed by pre-processing, lag-based feature preparation, chronological train-test splitting, model development, and performance evaluation. The predictive ability of the models was assessed using Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and the coefficient of determination (R). The results show a clear performance difference among the five models. Linear Regression produced the weakest results, with an MAE of 2.41C, MAPE of 7.12%, RMSE of 3.08C, and R of 0.837. KNN and SVR improved the prediction accuracy, while the ensemble models performed best overall. Random Forest achieved an MAE of 1.76C, MAPE of 5.21%, RMSE of 2.31C, and R of 0.912, whereas XGBoost outperformed all other models with the lowest error values and the highest fit, recording an MAE of 1.69C, MAPE of 4.96%, RMSE of 2.18C, and R of 0.928. These findings indicate that nonlinear ensemble-based methods are more effective than simpler linear and distance-based models for this prediction task. The study concludes that XGBoost provides the most reliable performance for temperature-based heatwave assessment and may serve as a practical model for short-term extreme heat analysis.