A demand response strategy for direct expansion air conditioning systems combining self-modeling and reinforcement learning

This study addresses the challenges faced by existing research in direct-expansion (DX) air conditioning demand response control, particularly high modeling costs and privacy concerns. A novel demand response control strategy is proposed, which integrates low-cost modeling with RL techniques. The self-learning algorithm leverages sensors, controllers, and IoT-based weather data to quickly learn the dynamic thermal and humidity characteristics of a room, based on a 2-h indoor temperature and humidity change. This self-learning model is subsequently used to design an intelligent control strategy for the DX air conditioning system, derived through cloud-based RL. A case study validates the effectiveness of the approach, demonstrating that the demand response control strategy, trained using Proximal Policy Optimization (PPO) with the self-learning dynamic thermal and humidity model, outperforms both single temperature control and energy-saving strategies. The results show significant reductions in energy consumption and electricity costs while improving responsiveness to time-of-use electricity pricing.