Machine learning based thermal comfort prediction in office spaces: Integrating SMOTE and SHAP methods

Accurate thermal comfort prediction is essential for enhancing indoor environmental quality and minimizing building energy use. The scarcity of individual thermal sensation data poses significant challenges in developing robust prediction methodologies. This study introduces an efficient model for predicting thermal comfort using limited datasets, utilizing the SMOTE technique for data enhancement and Bayesian optimization to improve machine learning efficacy. Field surveys collected environmental and questionnaire data from office spaces to train and optimize six machine learning models. The BO-XGBoost model excelled in predicting TSV, achieving an MAE of 0.1787, RMSE of 0.4343, R² of 0.8459, and accuracy of 85.81%. SHAP analysis identified air velocity, clothing thermal resistance, mean radiant temperature, and air temperature as critical determinants of thermal comfort. The model demonstrates high accuracy with limited data and could significantly enhance HVAC system optimization, thereby reducing energy consumption and carbon emissions, offering considerable practical advantages.