A Hybrid Machine Learning Framework for Operational 24-Hour-Ahead Forecasting of Human Heat Stress in Hyper-Arid Climates: A Case Study of Iraq

Extreme heat poses escalating risks to human health, labor productivity, and infrastructure resilience in hyper-arid regions. As climate change intensifies the frequency and severity of heatwaves, operational early warning systems with meaningful lead time are critical for strengthening climate resilience and supporting adaptation to extreme heat. In this study, we developed a hybrid machine learning (ML) framework for operational 24-hour-ahead forecasting of the Universal Thermal Climate Index (UTCI) in Iraq, one of the world’s most heat-vulnerable regions. The framework integrates a Twin Support Vector Machine (TSVM) optimized using a hybrid Whale Optimization Algorithm–Salp Swarm Algorithm (HWOA-SSA) to effectively capture the nonlinear and highly dynamic relationships among air temperature (AT), solar radiation (SR), relative humidity (RH), and wind speed (WS). The developed TSVM was validated against Support Vector Machine (SVM), Relevance Vector Machine (RVM), and Random Forest (RF) models. Using hourly ERA5 reanalysis data (1981–2020), the model was trained to generate high-resolution UTCI forecasts with a 24-hour lead time, and its performance was evaluated using a chronological calibration-validation split, in which the earlier portion of the time series (1981–2008) was used for model calibration and the later period (2009–2020) was reserved exclusively for validation. The optimized TSVM demonstrated strong predictive performance on temporally independent validation data, achieving a Kling–Gupta Efficiency (KGE) of 0.983, a correlation coefficient (R²) of 0.991, and a normalized RMSE of 0.206. The framework accurately reproduced diurnal amplification patterns and the spatial heterogeneity of extreme heat stress across Iraq, including regions that frequently exceed very strong heat stress thresholds during summer afternoons. By enabling reliable lead-time forecasting of hourly heat stress, the proposed framework offers a promising hybrid machine-learning approach for future heat early-warning applications and may be adapted to other hyper-arid and heat-prone regions through local calibration and validation.