Data-driven Thermal Comfort Prediction Analysis

This paper presents a data-driven thermal comfort predictor based on machine learning algorithms applied to the open-source RP-884 Thermal Comfort Database. The predictor combines multiple weak predictors through voting-based ensemble learning to achieve superior generalization capabilities. Ablation experiments are conducted on the input features using the pre-trained predictor to investigate their impact on model performance and significance. The results show that the data-driven predictor based on suitable input features can immensely reduce MAE and RMAE and attain an overall significance comparable to that of traditional regression models.