Enhancing cross-individual generalizability of thermal sensation prediction in dynamic environments: integrating physiological, emotional, and environmental features

Thermal perception prediction is essential for occupant-centered thermal environment control. However, current thermal perception models prioritize individual-level accuracy while neglecting cross-individual generalizability, limiting their applicability in spaces with frequently changing occupants. This study examined emotion–thermal coupling and developed both personalized and generalized thermal sensation prediction models using eight machine learning (ML) algorithms for known and unknown individuals. Notably, this study is the first to reveal significant sex differences in emotional dominance, with males exhibiting a stronger sense of control than females under cold and hot thermal conditions. Under dynamic thermal environments, thermal sensation showed weak correlations with physiological signals, whereas associations with emotional states remained robust. Predicting thermal sensation also proved more challenging in dynamic than in steady-state conditions, particularly for unseen individuals. Nevertheless, ambient temperature, thermal conditions considering the temperature change direction (TC-TCD), and emotional states significantly enhanced model cross-individual generalizability. Personalized model accuracies reached 94.97% for males and 96.49% for females, while generalized model accuracies improved from 58.75% to 81.31% for males and from 71.15% to 85.61% for females. Emotional states were key predictors of thermal sensation, primarily driven by valence, which exhibited a clear U-shaped relationship with thermal sensation votes (TSV). Finally, high generalizability was achievable with only five features, with generalized model accuracies decreasing by merely 0.67% for males and 0.59% for females. By integrating environmental and emotional information, this study significantly advances the cross-individual generalizability of thermal sensation prediction.