Adaptive multi-agent HVAC control for thermal comfort using multi-agent PPO with population-based training

Songze Guan

doi:10.1016/j.enbuild.2025.116882

Adaptive multi-agent HVAC control for thermal comfort using multi-agent PPO with population-based training

2025-12-25

0

PRE

AI

Sivanand Somasundaram

A

Alex Chong

C

Christopher H. T. Lee

C

Chau Yuen

DOI：10.1016/j.enbuild.2025.116882

Original

Original request for help

Save

Abstract

En 中文

Effective control of heating, ventilation, and air conditioning (HVAC) systems is critical not only for ensuring indoor thermal comfort, but also for reducing energy consumption and mitigating their significant contribution to global carbon emissions. Recent reinforcement learning approaches for HVAC control often struggle with issues such as unstable convergence, suboptimal coordination among zones, and sensitivity to hyperparameter settings, limiting their real-world applicability. To address these challenges, this study proposes a novel control algorithm combining Population Based Training (PBT) and Multi-Agent Proximal Policy Optimization (MAPPO). The proposed method MAPPO-PBT automatically tunes hyperparameters and trains the MAPPO model to enable adaptive adjustment of cooling setpoints, thereby optimizing energy efficiency and thermal comfort in a coordinated manner. The effectiveness of the proposed approach is validated through EnergyPlus simulations using a medium-sized two-story office building modeled in DesignBuilder. Compared to traditional rule-based control method, MAPPO-PBT achieves much better thermal comfort, represented by the Predicted Mean Vote (PMV), showing a 28.2% reduction compared to rule-based control. Furthermore, MAPPO-PBT maintains average room PMV within 0.2 and ensures setpoint stability comparable to rule-based strategies.

Adaptive multi-agent HVAC control for thermal comfort using multi-agent PPO with population-based training

Abstract

Journal

Organization