Temperature-preference learning with neural networks for occupant-centric building indoor climate controls

Abstract Heating, ventilation, and air-conditioning (HVAC) are vital components in providing a comfortable indoor climate for the occupants of buildings. In commercial buildings, HVAC setpoints are set according to average comfort temperatures. However, individual temperature preferences may be different. The purpose of this study is to explore the means of making HVAC systems respond automatically to local occupant temperature preferences. To create an occupant-centric indoor temperature environment, we propose an online-learning-based control strategy together with its design process. Four essential variables from four domains—time, indoor and outdoor climates, and occupant behavior—are extracted to construct datasets for preference models. A neural network algorithm and corresponding hyperparameters are suggested to model temperature preferences. According to time-dependent setpoints learned from dynamic contexts, a set of specified rules is used to determine setpoints for HVAC systems. For a period of five months, the resulting learning-based temperature preference control (LTPC) was applied to a cooling system of an office space under real-world conditions. The four case study rooms consisted of typical office uses: single-person and multi-person offices. The experimental results indicate that occupant preferences in the individual rooms differ from each other in both time horizon and temperature levels. The results report energy savings of between 4% and 25% as compared to static temperature setpoints at the low values of preferred temperature ranges. Meanwhile, during LPTC, the need for occupant interventions for adjusting room temperatures to fit their preferences was reduced from four to nine weekdays a month to a maximum of one weekday a month.