Interpretable prediction of thermal sensation for elderly people based on data sampling, machine learning and SHapley Additive exPlanations (SHAP)

Abstract

Machine learning (ML) algorithms are frequently used to predict human thermal sensation votes (TSV). Establishing a TSV prediction model for elderly people is essential for improving thermal comfort and ensuring physiological health. This paper aims to combine ML algorithms with data sampling methods to establish a TSV prediction model for elderly people and provide an interpretation of the model based on the SHapley Additive exPlanations (SHAP) method. Firstly, 44 elderly people from 2 pensioners’ buildings are recruited as the participants, and the summer environmental parameters, physiological parameters and TSV are collected. Then, 7 ML algorithms and 8 data sampling methods are used to predict the 3-point TSV. Finally, the importance ranking, the positive or negative effects and the interaction of the features are analyzed based on the SHAP method. The results indicate that, the Tomek Links + Synthetic Minority Over Sampling Technique + Xgboost model performs the best. The F1 scores of “cool”, “neutral” and “warm” are 73%, 79% and 72%, respectively. Air temperature (TA), mean skin temperature (MST), body mass index (BMI) and relative humidity (RH) are the four most important features. For elderly people in summer, the indoor thermal neutral TA, RH and MST are about 29 °C, 45% and 36 °C, respectively. This paper can be adopted to provide method support for predicting the TSV of elderly people and provide data reference for the indoor environmental parameters of the pensioners’ buildings.