Predicting thermal sensation vote in an educational building: A comparative assessment of PMV-based mapping and machine learning approaches

Ensuring thermally comfortable indoor conditions is a fundamental requirement for occupant health, productivity, and overall wellbeing. However, despite advances in modeling techniques, predicting thermal comfort using experimentally collected real-world classroom data remains insufficiently explored and requires further investigation. This study presents a comparative analysis between two approaches for predicting Thermal Sensation Vote (TSV) in an air-conditioned educational building during the cold season. In the first approach, the Predicted Mean Vote (PMV) was calculated using the Fanger model and mapped to TSV classes using the discretization criteria. In the second approach, four ensemble machine learning classifiers including Random Forest (RF), Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGBoost), and LightGBM were trained on data collected from the occupants. A two-stage feature selection framework combining Spearman rank correlation and Recursive Feature Elimination (RFE) was applied prior to model training, and Bayesian optimization was employed to tune the hyperparameters of each classifier. The PMV-based approach achieved an overall accuracy of 27.9%, while the best-performing machine learning model which is tuned by Bayesian optimization reached an accuracy of 82.17%. This marked an improvement of over 54 percentage points compared to the PMV benchmark. The results show that all four optimized machine learning models outperformed the PMV-based estimation in predicting TSV.