Explainable machine-learning surrogates for sun-patch-driven indoor thermal comfort

Localized solar patches generate transient indoor thermal asymmetries that conventional steady-state models cannot adequately capture, particularly in environments conditioned by radiant floor systems. Under these circumstances, the heated slab is subjected to additional radiative loading, which can exacerbate thermal comfort risks by inducing localized overheating. This study develops interpretable data-driven surrogate models to predict the short-term evolution of global and local thermal comfort in an experimental 1:1 test chamber equipped with a hydronic floor system. Using a ‘delta-prediction’ strategy to capture system dynamics, seven machine learning algorithms are trained to predict Body Average Comfort (BAC), Body Average Sensation (BAS), and local temperatures near the head and ankle over a 10-minute horizon. Ensemble tree-based and kernel-based models regression generally outperformed linear baselines, achieving high generalization on an independent test period: R² reached 0.97 for ankle temperature (RMSE≈0.32℃), 0.93 for BAS, and 0.85 for BAC. Head temperature proved the most challenging target (R²≈0.81, RMSE≈1.36℃) due to complex radiative interactions. Grouped SHAP analysis reveals distinct physical drivers: while global comfort is dominated by operative temperature and humidity, direct solar radiation is the overwhelming driver of local temperatures and a significant contributor to thermal sensation. A counterfactual Sun Patch Impact (SPI) analysis quantifies this effect, showing that solar exposure systematically elevates predicted sensation by up to 0.05 scale units at peak irradiance (800 W/m²). Finally, a principal component analysis identifies a ‘local–global mismatch’ phenomenon, where 18–20% of globally neutral states conceal local overheating risks at the ankle level, highlighting the necessity of multi-objective monitoring for effective sun-patch management.