Multi-agent distributed reinforcement learning for energy-efficient thermal comfort control in multi-zone buildings with diverse occupancy patterns

The rapid development of smart cities and automated infrastructures has increased building electricity demand, particularly from heating, ventilation and air conditioning (HVAC) systems. Current HVAC control methods primarily address short-term dynamics and single-zone scenarios, overlooking complexities from seasonal variability and diverse occupancy patterns in multizone buildings. Furthermore, existing data-driven frameworks lack mechanisms to transfer control policies across buildings with different thermal zone configurations. To address these limitations, this study proposes a decentralized multi-agent reinforcement learning framework for energy-efficient thermal comfort management in multizone buildings. Transfer reinforcement learning enables efficient adaptation of control strategies to buildings with differing zone configurations. Results demonstrate that occupancy and zone-specific control actions effectively balance energy efficiency and occupant comfort. The proposed method maintains thermal comfort within acceptable levels while reducing grid energy import by 51.7 % compared to conventional rule-based methods. Assigning a higher energy weight in the decentralized network structure achieved an additional 23 % reduction in energy use. The transfer learning approach successfully adapted control policies from a nine-zone office to a five-zone residential building with limited monitoring data and reduced building load by 6.4 %. Practically, this approach significantly reduces training data requirements and accelerates model deployment. Collectively, these enhancements provide building operators with effective tools to achieve significant energy savings and support city-level sustainability efforts.