Topology-aware hypergraph reinforcement learning for indoor occupant-centric HVAC control

Dianyu Zhong, Kailai Sun, Tian Xing, Ziyou Zhang, Qianchuan Zhao, Jian Kang

Abstract: Heating, ventilation, and air conditioning (HVAC) systems account for a significant portion of global energy consumption, posing a significant obstacle to achieving Net Zero emissions by 2050. Existing reinforcement learning (RL) based occupant-centric control strategies show promise but face limitations in real-world settings due to insufficient integration of occupant activity data and underutilization of building spatial structures. To address these challenges, this study introduces a novel framework that integrates real-time, vision-based occupant activity recognition and models the building’s spatial topology as a hypergraph, enabling topology-aware value decomposition. Our approach outperforms benchmark RL algorithms, achieving state-of-the-art performance. Experiments based on real-world office data demonstrate that integrating occupant activity reduces the predicted percentage of dissatisfied by 20.9 % and improves energy efficiency by 21.1 %; separately, leveraging building topology yields a 25.6 % reduction and a 13.7 % efficiency gain. These findings offer new insights into intelligent control for energy-efficient, occupant-centric buildings and confirm the framework’s potential for large-scale, real-world deployment.