Comparative analysis of thermal preference prediction performance in different conditions using ensemble learning models based on ASHRAE Comfort Database II

Prediction of thermal comfort of building occupants using ensemble learning models is a hot research topic. The performance of ensemble learning models used to predict thermal preference may not only be determined by its algorithm structure, but also by the parameters of the data set chosen for training the learning models as well as the characteristics (building type and season) of the dataset. In this paper, based on precision, recall, F1-score, weighted F1-score, the prediction performance of 10 machine learning models (6 traditional and 4 ensemble models) trained with different data subsets was compared systematically, and the characteristics of ASHRAE Comfort Database II were used for the first time to observe the performance of ensemble learning models. The feature importance of the variables was calculated by random forest (RF) and gradient boosting decision tree (GBDT) to explore the key parameters influencing thermal preference prediction. The results showed that the performance of the ensemble models achieved the greatest improvement in the process of training data increasing from 40% to 60%. After training with the data from the classroom during the summer, the ensemble learning models showed a significant performance based on the weighted F1-score. Furthermore, compared with other models, RF and deep cascade forest (DCF) showed significant advantages in predicting thermal preference with different data subsets. Therefore, RF and DCF with selected key parameters of thermal preference can be used to predict individual thermal preference in different conditions, providing references for automatic regu-lation of building thermal environments.