Machine learning modeling of indoor thermal sensation under solar radiation considering skin temperatures

The increasing adoption of transparent building envelopes has made it essential to consider the impacts of solar radiation on indoor thermal sensation. Traditional regression models fall short of predicting thermal sensation when confronted with a multitude of input parameters and intricate interactions. Notably, prevailing prediction models for skin temperatures, one of crucial input parameters, often fail to adequately account for the effects of solar radiation. Therefore, this study modifies the JOS-3 model for skin temperature prediction and validates its reliability through indoor thermal sensation experiments. Additionally, machine learning (ML) techniques which are well-suited for capturing complex relationships among parameters are adopted to predict thermal sensation under solar radiation, then the prediction performances of four ML algorithms are compared. Results show that the modified JOS-3 model improves prediction accuracy by over 50 %, with little mean errors between predicted and measured skin temperatures for most body parts. In ML modeling, the average cross-validation accuracy can be enhanced by 20 % when incorporating skin temperatures as input features, and combining multiple types of features leads to better prediction performance than that with a single type. Furthermore, the differences in local skin temperatures between symmetric body parts, resulting from solar radiation asymmetry, are the key factors influencing thermal sensation. Four ML algorithms all achieve prediction accuracies over 0.74 for thermal sensation, and they are more effective at identifying neutral and very hot thermal sensations of occupants among all TSV categories.