Dimension analysis of subjective thermal comfort metrics based on ASHRAE Global Thermal Comfort Database using machine learning

We analyzed the ASHRAE Global Thermal Comfort Database II to answer a fundamental but overlooked question in thermal comfort studies: how many and which subjective metrics should be used for the assessment of the occupants' thermal experience. We found that the thermal sensation is the most frequently used metrics in Thermal Comfort Database II, followed by thermal preference, comfort and acceptability. The thermal sensation/thermal preference, thermal comfort/air movement acceptability and thermal comfort/thermal preference are the top three most dependent metrics pairs. A principal component analysis confirmed that the personal experience of thermal conditions in built environment is not a one-dimensional problem, but at least a two-dimensional problem, and suggested thermal sensation and thermal comfort should be asked in right-now surveys as the first two Principal Component are majorly constructed by thermal sensation and thermal comfort. To further confirm the predictive power of thermal sensation and comfort, we used logistic regression and support vector machine to predict thermal acceptability and thermal preference with thermal sensation and comfort. The prediction accuracy is 87% for thermal acceptability and 64% for thermal preference. The prediction error might be due to occupants' individual difference and people errors in answering survey. These findings could help the design of chamber experiments, field studies, and human-building interaction interfaces by shedding light on the choice of subjective thermal metrics to effectively and accurately collect information on occupants’ thermal experience.