Machine learning for multi-zone thermal comfort prediction in winter indoor sports venues

To address the highly dynamic thermal environments of large-span winter sports venues and the limited applicability of traditional thermal comfort models, this study constructed multi-zone thermal comfort prediction models using machine learning, with typical badminton and basketball gymnasiums in regions with hot summers and cold winters as research subjects. Field measurements were conducted to collect environmental parameters, including air temperature, humidity, air speed and mean radiant temperature, while subjective assessments employed a three-dimensional thermal perception framework comprising thermal sensation vote, psychological adaptation value and environmental expectation variance. These integrated objective-subjective datasets enabled machine learning modelling analysis using random forest, XGBoost and other algorithms, revealing that the metabolic compensation effect induced by moderate-intensity exercise effectively offsets heat loss in low-temperature environments, thereby maintaining a neutral thermal sensation. The time variable, serving as a proxy for environmental dynamics, exhibited significantly higher feature importance compared to traditional static parameters in thermal comfort prediction. Data-driven models demonstrated significant advantages over the predicted mean vote index and effectively captured nonlinear interactions amongst variables. The study demonstrated that, in shallow temperature gradients, large-scale spatial scenarios, the synergistic interaction between temporal behavioural characteristics and environmental dynamic responses primarily governed thermal comfort variations observed.