Non-intrusive Prediction of Indoor Occupant Thermal Comfort Via Infrared Thermography and Machine Learning

The indoor occupant thermal comfort prediction model is essential for the development of personalized environmental control systems, which helps to balance human thermal comfort and building energy consumption. Non-invasive measurement of facial temperature is an important data source for developing personal thermal comfort models. However, previous studies have always relied on a combination of visible and infrared images, making them unsuitable for application in low-light environments. To address this limitation, this study adopted YOLOv8 and Dlib to identify the temperature features of facial ROIs on infrared images directly. Machine learning algorithms (i.e., RF, SVM and KNN) were then applied to predict subjective thermal comfort responses. The results showed that the performance of prediction model based on KNN outperforms RF and SVM, achieving a prediction accuracy over 78.57% for thermal comfort indicators, which represents an accuracy improvement of over 13% compared to the PMV model.