Machine-learning-based personal thermal comfort modeling for heat recovery using environmental parameters

Thermal comfort modeling based on average response of occupants has been of interest in building performance assessment for many years. Energy consumption optimization was the main concern initially, but recently, it was shown that poor health was significantly associated with thermal discomfort. It is crucial to control hospitals' thermal comfort requirements, particularly when the patient and the caregiver have to stay in the same room for a long time. These goals may be obtained if personal comfort models, that predict individuals' comfort responses, are considered instead. The primary goal in this study was to investigate the effects of physiological and environmental factors on individuals' thermal preferences in indoor environments using datasets collected for the Green Buildings Innovation Cluster (GBIC) project. Hence, several feature sets were created considering various combinations of the physiological, environmental and personal factors and several machine learning classification algorithms were utilized to develop personal comfort models, including Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbor (KNN) and an ensemble algorithm in two different approaches: multi classification and binary classification. Models' Performances were evaluated and compared employing several metrics. The median accuracies of 0.89 and 0.84 were obtained considering the models with the best performances on the cross-validation sets in multi-classification and binary classification approaches, respectively. Furthermore, more complex models such as RF and SVM predicted individuals' thermal preferences more precisely yet required more computational time than other classifiers. The results in this study indicated that skin temperature is a strong predictor of personal thermal preferences and a combination of several physiological parameters, including metabolic rate and heart rate, with the skin temperature data as inputs for the models would provide higher predicting accuracies.