Bayesian robust reinforcement learning for coordinated air conditioning and energy storage system control in high-performance residential buildings under forecast uncertainty

In high-performance residential buildings, centralized air conditioning using a single unit is commonly adopted to improve energy efficiency under low load conditions. However, this strategy often results in frequent defrosting during winter, reducing thermal comfort and increasing electricity consumption. Although reinforcement learning shows promise for building energy control, especially when incorporating weather and electricity price forecasts into the state, its performance tends to deteriorate significantly under prediction errors. To address this issue, this study develops a Bayesian robust RL method for the joint control of air conditioning and battery systems. A physics-driven defrost evaluation module is integrated to dynamically estimate heating performance under frosting conditions. During training, structured perturbations constructed from prior knowledge are introduced to emulate realistic forecast errors, and a Kullback–Leibler (KL) divergence-based regularization term is added to the policy objective to reduce sensitivity to input disturbances. These mechanisms enable the control strategy to proactively identify frosting risks and preemptively lower the air conditioner setpoint, leveraging the building’s thermal storage to suppress defrosting events on the user side—without requiring hardware modifications. Simulation results under three forecast scenarios, ranging from perfect accuracy to high uncertainty, demonstrate that the proposed method effectively maintains thermal comfort while significantly reducing defrost duration and electricity costs. Notably, under severe forecast errors, the method achieves up to an 8.2 % reduction in electricity costs compared to the baseline, confirming its robustness and practical applicability in real-world building energy management systems.