Performances of machine learning algorithms for individual thermal comfort prediction based on data from professional and practical settings

Abstract

Individual thermal comfort prediction models based on real-time monitoring parameters could improve the efficiency of personal air conditioning systems. However, there is a contradiction between accuracy and cost/convenience to acquire input data for individual thermal comfort prediction models. In previous studies, the performance of these models was examined using input data from either professional or low-cost devices, but they have not been compared with each other. In this study, 32 subjects with different individual features participated in climate chamber experiments with air temperatures ramp between 18 and 34 °C in summer. Commonly-used machine learning (ML) algorithms, e.g. Support Vector Machine, Decision Tree, K-Nearest Neighbors (KNN), Discriminant Analysis and ensemble methods, were used to examine two kinds of input data: the professional setting with 10 predictors include multiple personal features and high accuracy skin temperatures (±0.15 °C), and the practical setting with 5 easier obtained predictors include infrared skin temperatures (±1 °C). Results showed the overall accuracies of ML algorithms were generally higher by 0.4%–12.3% using the data in professional setting than that in practical setting. Comprehensive consideration, the optimized Cosine KNN (an accuracy of 83.6%, a precision of 89.7%, a recall of 84.3% and AUC = 0.86) and ensemble of Subspace KNN (an accuracy of 75.4%, a precision of 79.6%, a recall of 82.7%, and AUC = 0.87) were the best ML algorithms for professional and practical settings, respectively. The sensitivity analysis showed the skin temperatures were important predictors, while the age was a more important feature in practical setting then professional setting. This study could provide useful information for the selection of features and machine learning algorithms for individual thermal comfort prediction in actual buildings.