Personal Thermal Comfort Modeling by Machine-Learning

Thermal comfort modeling based on an average response of a group of occupants has been applicable in building performance assessment for many years. Energy consumption optimization was initially the primary concern, but recently, it was shown that poor health was significantly associated with thermal discomfort in many cases. Optimized energy consumption in a building and a healthy thermal comfort situation in hospitals may be obtained if personal comfort models are considered instead. In this paper, our primary goal is to investigate the effects of physiological and environmental factors on occupants thermal comfort, and datasets collected for the GBIC project are used. We utilized several machine learning classification algorithms to develop personal comfort models, including Random Forest, Support Vector Machine, K-Nearest Neighbor, and an ensemble algorithm in a multi-classification approach. The accuracy of each model was compared. It was shown that the personal comfort models based on the multi-classification approach resulted in a median accuracy of 0.89, considering the best models trained on all of the available features. The result of our study suggests that skin temperature is a strong predictor of personal thermal preferences, and considering a combination of several physiological parameters such as metabolic rate and heart rate with the skin temperature as input for the models would result in higher accuracies.