Reinforcement Learning Based Optimization Method for Comfort-Energy Balance in Multi-Zone HVAC System

The operation of heating, ventilation, and air conditioning (HVAC) systems in modern buildings often suffers due to the limited adaptability and efficiency of conventional rule-based control mechanisms. Addressing the need for more rational and flexible operation, we introduce a novel multi-objective optimization approach leveraging Deep Q-Network (DQN) for open office environments. Our method aims to strike an optimal balance between minimizing energy consumption and maximizing occupant comfort within the air conditioning system. Unlike rule-based methods, our data-driven approach circumvents the need for extensive a priori knowledge of building-specific thermodynamics, thereby enhancing adaptability and ease of deployment. We developed a building simulation model for the target open office area, complemented by a custom-built simulation and testing platform integrated with a building modeling interface. This platform facilitates real-time data acquisition and parameter adjustment, enabling the DQN agent to efficiently control HVAC operations. By dynamically adjusting temperature setpoints across various thermal zones via variable air volume (VAV) actuators, the DQN agent skillfully navigates the energy-occupant comfort trade-off. Empirical evidence from extensive training iterations indicates that our DQN-based control method outperforms traditional rule-based systems, reducing energy consumption by up to 24% while maintaining optimal occupant comfort. The experimental results demonstrate the efficacy of the proposed method for energy optimization in building management.