Personalized thermal satisfaction prediction for the elderly using machine learning with SHAP interpretation

The predicted mean vote (PMV) model is overly generalized and shows limited applicability to older adults. This study develops a personalized machine learning (ML) model based on the XGBoost algorithm to improve thermal satisfaction prediction and support personalized indoor environmental control. A longitudinal field study was conducted in an elderly care center during the heating season from 2023 to 2024. A total of 1008 samples were collected, of which 900 were retained after data cleaning. Irrelevant and redundant features were removed, and the best-performing model was determined through data balancing and hyperparameter tuning. Model interpretation was performed using SHapley Additive exPlanations (SHAP). The ML model achieved an accuracy of 0.909, outperforming the PMV model, indicating improved precision in predicting thermal satisfaction among older adults. The F1 score of 0.79 demonstrates that the model effectively balances precision and recall, successfully identifying both satisfaction and dissatisfaction cases. These results support enhanced safety and personalized thermal comfort for older adults and contribute to more refined indoor environmental management in elderly care settings.