Using machine learning algorithms to predict occupants’ thermal comfort in naturally ventilated residential buildings

Thermal comfort evaluations in the built environment are essential to occupant's satisfaction and also building energy consumption. Traditionally, thermal comfort has been assessed by Fanger's PMV model that was developed based on extensive laboratory tests. However, it's been found that PMV predictions are not good in field studies, especially in buildings with natural ventilation. In this paper, machine learning (ML) algorithms were used to predicted occupants’ thermal comfort votes (TCV) and thermal sensation votes (TSV), using 5512 sets of thermal comfort data collected in naturally ventilated residential buildings in fourteen cities in China. Environmental parameters, personal parameters, climatic types, and adaptive control measures were considered and used as input parameters for the ML model. It was found that environmental parameters (both outdoors and indoors), personal parameters (metabolic rate and clothing insulation), and climatic types all significant affect both TCV and TSV, while adaptive control measures only affect TSV but not TCV. Comparing with established models (PMV, ePMV and aPMV), the ML models had smaller errors in predicting TSV and TCV. Based on these results, we suggest that the ML, especially ANNs model was reliable in predicting occupants’ TCV and TSV in naturally ventilated residential buildings, and performed better than traditional thermal balance based models. Moreover, using the new model, we found that acceptable temperature ranges in naturally ventilated buildings were far wider than the ASHRAE adaptive comfort zone, suggesting that appropriate models should be established based on local data.