Multi-agent deep reinforcement learning based HVAC control for multi-zone buildings considering zone-energy-allocation optimization

• Published in: Energy and buildings 2025-02, Vol.329, p.115241, Article 115241
• Main Authors: Xue, Wenping, Jia, Ning, Zhao, Mengtao
• Format: Article
• Language: English
• Subjects: Building energy saving; Genetic algorithm; Multi-agent deep reinforcement learning; Multi-zone HVAC control; Thermal comfort; Zone energy allocation
• ISSN: 0378-7788
• DOI: 10.1016/j.enbuild.2024.115241

Advanced control of the heating, ventilation and air conditioning (HVAC) system aims to minimize the HVAC energy consumption while guaranteeing the thermal comfort of occupants. Through interactions with the building environment, deep reinforcement learning (DRL) enables the continuous or high-dimensional optimal HVAC control without knowing the building thermal dynamics model. This study proposes a hybrid method for multi-zone HVAC control based on combination of the genetic algorithm (GA) and multi-agent deep deterministic policy gradient (MADDPG) algorithm. Different from the commonly used reward design for zone agent, the form of weight coefficients is adopted to allocate the energy consumption to each zone. The GA is utilized to optimize these weight coefficients for the training of the MADDPG algorithm. Performance of the proposed GA-MADDPG control method is validated through simulation in TRNSYS. Results demonstrate that, compared with the fixed setpoint control (26∘C), model predictive control (MPC), multi-agent deep Q-network (MADQN) and MADDPG methods, the proposed method can respectively reduce the HVAC energy consumption by 6.7%, 2.9%, 4.0% and 2.0%, as well as maintain the thermal comfort of zone occupants. •Zone-energy-allocation optimization is considered in the agent reward design.•A hybrid GA-MADDPG control method is proposed for multi-zone buildings.•The proposed method is validated through comprehensive comparisons.