Predictive Modeling of Urban Heat Stress Using Machine Learning: A 2050 UTCI-Based Assessment

Abstract

Urban heat islands and increasingly severe thermal stress pose critical challenges to sustainable urban development, yet physics-based simulation tools remain computationally expensive and ill-suited to early-stage climate-responsive design. This study applies machine learning to predict outdoor air and surface temperature using 20 years (2005-2025) of daily maximum temperature data ranging from 15°C to 40°C. Four algorithms were evaluated: Linear Regression, Random Forest, Gradient Boosting and Extreme Gradient Boosting (XG Boost). Gradient Boosting achieved the best performance with RMSE = 1.869, MAE = 1.405, and R² = 0.708, indicating robust predictive ability across diurnal temperature and seasonal patterns. 2050 projections, based on direct extrapolation of the trained model's seasonal temperature pattern, show a shift toward moderate-to-strong Universal Thermal Climate Index (UTCI) heat stress conditions across the majority of the year, underscoring the need for climate-responsive urban design even under conservative, non-scenario-scaled projections. The results strengthen the acceptance of data-driven ML frameworks for the application in urban planning and climate-responsive architectural design practice.