Machine learning and features for the prediction of thermal sensation and comfort using data from field surveys in Cyprus.

Perception can influence individuals' behaviour and attitude affecting responses and compliance to precautionary measures. This study aims to investigate the performance of methods for thermal sensation and comfort prediction. Four machine learning algorithms (MLA), artificial neural networks, random forest (RF), support vector machines, and linear discriminant analysis were examined and compared to the physiologically equivalent temperature (PET). Data were collected in field surveys conducted in outdoor sites in Cyprus. The seven- and nine-point assessment scales of thermal sensation and a two-point scale of thermal comfort were considered. The models of MLA included meteorological and physiological features. The results indicate RF as the best MLA applied to the data. All MLA outperformed PET. For thermal sensation, the lowest prediction error (1.32 points) and the highest accuracy (30%) were found in the seven-point scale for the feature vector consisting of air temperature, relative humidity, wind speed, grey globe temperature, clothing insulation, activity, age, sex, and body mass index. The accuracy increased to 63.8% when considering prediction with at most one-point difference from the correct thermal sensation category. The best performed feature vector for thermal sensation also produced one of the best models for thermal comfort yielding an accuracy of 71% and an F-score of 0.81.