Deploying deep reinforcement learning for low-level HVAC control in multi-zone buildings: A comparative study with ASHRAE G36 sequences

This paper proposes a methodology for optimizing HVAC control in multi-zone buildings using Deep Reinforce- ment Learning. The study focuses on optimizing the central AHU system by controlling all low-level components within both the air and water loops, addressing the complex dynamics of multi-zone interactions. The case study is based on a building within the Politecnico di Torino campus. Modelica-based simulations are used to model both the HVAC system and building dynamics, allowing the integration and evaluation of the ASHRAE G36 con- trol standard as a benchmark. Two DRL strategies are developed and evaluated, Zone-Aware and Zone-Integrated, under both winter and summer conditions, with the goal of improving energy efficiency, indoor temperature con- trol, and indoor CO2 concentration, under varying occupancy profiles. The results reveal that both DRL strategies outperform the G36 baseline in terms of energy savings (up to 17 %), indoor temperature violations, and CO2 concentration. Additionally, DRL controllers demonstrate strong generalizability and adapt seamlessly to unseen occupancy profiles without manual tuning. This research highlights the potential of DRL to provide scalable, adaptive, and energy-efficient HVAC control solutions for multi-zone buildings.