Interpretable thermal comfort prediction based on SHAP-KANs: Revealing seasonal differences in indoor-outdoor factor influences

Thermal comfort research has documented differential indoor and outdoor impacts on human thermal perception, yet the mechanisms driving these differences across seasons remain unclear. This study aims to predict thermal sensation votes and to reveal how environmental and behavioral factors differ between indoor and outdoor settings across seasons. We propose a machine learning framework combining SMOTE for data balancing, Bayesian optimization for tuning, and SHAP for interpretability, comparing Random Forest, XGBoost, and KANs using environmental parameters, clothing insulation, metabolic rate, and Hunger Voting. Field data were collected from indoor and outdoor environments in Wuhan across three seasons. The KANs model outperformed the other algorithms, achieving accuracies of 0.9074 (indoor transitional), 0.8421 (outdoor summer), and 0.8889 (outdoor winter low-temperature). SHAP analysis reveals distinct seasonal and spatial patterns: winter indoor comfort is driven by HV and clo, while outdoor comfort depends on wind speed, metabolic rate, and temperature; summer humidity shows opposite effects indoors versus outdoors; transitional seasons display a balanced interplay of multiple factors. These findings validate seasonal and spatial differences in thermal comfort mechanisms and offer insights for environmental design.