Energy-Efficient HVAC Control based on Reinforcement Learning and Transfer Learning in a Residential Building

The energy consumption of Heating, Ventilating, and Air conditioning (HVAC) systems is one of the most sig-nificant part of building energy consumption. In recent years, with the development of Reinforcement Learning (RL), many approaches has been proposed for building operation and energy conservation. However, previous work failed to establish accurate environments for RL, leading to the limited deployment of RL systems in real-world buildings. This paper proposes a data-driven approach to construct an RL environment and utilizes transfer learning to achieve real-world building control for energy conservation. We conducted a six-day field test in a experimental nearly-zero residential building in Beijing. The HVAC for the residential building is an air-source heat pump with a fresh air system that uni-formly controls the air temperature of three bedrooms. The RL environment consists of multiple XGBoost models to predict temperature and energy consumption, and a Deep Deterministic Policy Gradient (DDPG) agent which is pretrained on historical data of years 2020 and 2021, and applied to the real-world setting in 2023. The results indicate that our RL-based controller can effectively conserve energy in the experimental environment while maintaining thermal comfort, achieving a 11.33 % energy saving rate.