Calibrating thermal sensation vote scales for different short-term thermal histories using ensemble learning: Fid-Bau Portal

Calibrating thermal sensation vote scales for different short-term thermal histories using ensemble learning

Yuan, Liang / Qu, Rong / Chen, Tianyu / An, Na / Huang, Chenyu / Yao, Jiawei

Abstract The urban heat island effect intensifies, leading to increased thermal exposure for city residents. Variations in thermal sensation are observed among individuals with different short-term thermal experiences, challenging the reliability of Thermal Sensation Vote (TSV) scales. This study investigates TSV disparities in Shanghai, China, between two groups: individuals with Short-Term Air Conditioning Usage Experience (STACUE) and those Lacking Short-Term Air Conditioning Usage Experience (LSTACUE). Through questionnaire surveys and environmental monitoring, the study evaluates the performances of various ensemble learning models for predicting TSV. The selected Random Forest Regressor is employed to calibrate TSV scales for Physiologically Equivalent Temperature, Standard Effective Temperature (SET*), Universal Thermal Climate Index, and Perceived Temperature. Using Shapley values from game theory, we reveal how environmental variables contribute to TSV differences among individuals with varying short-term thermal histories. Results indicate higher Thermal Unacceptable Vote and greater Environmental Expectation Vote for STACUE individuals versus LSTACUE ones. Calibrated TSV scales, particularly SET*, exhibit significant enhancements over original scales: a 35.04 % increase in prediction accuracy percentage and a 0.57 correlation increase. Explainable analysis underscores that air temperature (28.8%) has a stronger impact on TSV among STACUE individuals, whereas mean radiant temperature (32.2%) is the primary factor affecting TSV among LSTACUE. Furthermore, we found gender interacts with thermal environmental parameters concerning TSV. This study sheds light on how short-term thermal history influences TSV among residents, informing customized urban thermal management strategies.

Highlights • Short-term thermal exposure history impacts the population's TSV. • The predictive performance of SET* utilizing ensemble learning is optimal. • The highest prediction accuracy percentage was elevated to 81.23 % in this study. • Attention is on the influence from Ta, Tmrt, gender and thermal interaction. • TSV displays dynamic nonlinear changes within populations.