Data-driven thermal preference prediction model with embodied air-conditioning sensors and historical usage behaviors

摘要

Predicting building occupants' thermal comfort, especially distinguishing individual thermal preference, is challenging but helpful for developing intelligent air conditioning (AC) control technology. Many studies have tried to solve this issue but heavily relied on extra sensing technology. This study sought to understand how occupants interact with AC devices and to predict occupants' thermal preference with limited sensors. Data from embodied sensors in 251 AC devices and occupants' interactions with these devices were analyzed. Five machine learning (ML) algorithms were applied to predict AC device's setting temperature changing actions. The results show that different occupants' AC usage behavior varied greatly in setting temperature preference and adjusting time. Users can be categorized as "prefer cool, " "prefer warm ", and "prefer neutral " according to the room temperature and setting temperature distributions; or as "regular type " and "irregular type " according to the time when the setting temperature was adjusted. More than 60% of users tended to set AC temperature in the range of 25C-28C. By applying random forest algorithm and proper data preprocesses, models can predict "increase setting temperature " and "decrease setting temperature " actions with 72.1%-87.3% accuracy. The model per-formance increases with larger samples and by adding Month and Hour as input features. With 30-50 times of training, the thermal preference learning curves for individual AC devices can reach relatively stable state. Lastly, a setting temperature control logic for air conditioners was discussed. Hopefully, this work can help to develop intelligent AC control methods that maximize occupants' thermal comfort while reducing energy consumption.