Physics-Informed Ensemble Learning for city-center grid cell temperature prediction during thermal extremes

Accurate prediction of city-region temperature is crucial for urban sustainability, which impacts energy management and public health. Current models struggle with spatio-temporal non-linearities and exhibit poor performance during extreme events due to data scarcity. We propose Physics-Informed Ensemble Learning (PIEL-NET). This novel hierarchical framework integrates a physics-terrain fusion agent with a regime-specialized agent trained under Regime Adaptive Focal Loss (RAFL). This architecture enables dynamic focus on underrepresented extreme cases where physical models underperform, distinguishing it from conventional physics-informed or ensemble methods. Our model processes the urban core and eight surrounding grid cells at 100 km resolution to capture advection context. Evaluated on decade-long hourly data from four global cities, PIEL-NET achieves a 69.2% error reduction below −7.4 °C in extreme cold conditions and attains a 0.89 Dice score for extreme event warnings, significantly outperforming five state-of-the-art benchmarks.