Predicting Human Thermal Comfort During Winter Heating Using Multi-Class Machine Learning Algorithms

To address the critical need for accurate human thermal comfort prediction in winter heating environments, this study established a comprehensive thermal comfort dataset containing 2089 valid samples through experiments. On this basis, thermal comfort prediction models were constructed using three multi-class machine learning algorithms: Support Vector Classification, K-Nearest Neighbors, and Random Forest. The predictive performance of 63 different feature combinations was systematically evaluated. The results indicate that the feature subset comprising indoor air temperature, forehead temperature, cheek temperature, dorsal hand temperature, heart rate, and systolic blood pressure yields the optimal prediction performance. Among the evaluated models, the Random Forest model demonstrated superior overall performance, achieving an accuracy exceeding 90% and an AUC ranging from 96% to 99%, significantly outperforming the SVC and KNN models. Compared with the traditional Predicted Mean Vote (PMV) model, the machine learning models developed in this study showed a substantial improvement in prediction accuracy under identical conditions; notably, the Random Forest model improved accuracy by approximately 40% over the PMV model. Based on these findings, a smart heating system framework integrating environmental sensors, wearable devices, and intelligent control valves is proposed, providing a theoretical basis and technical approach for realizing personalized and energy-efficient heating control.