A two-layer safe deep reinforcement learning framework for coordinated control of radiative cooling envelopes and HVAC systems

Integrating passive daytime radiative cooling (PDRC) envelopes into buildings is an effective strategy for reducing cooling energy demand. However, due to their limited cooling capacity, HVAC systems remain necessary to maintain indoor thermal comfort. The performance of PDRC envelopes is strongly influenced by meteorological conditions, limiting the effectiveness of conventional rule-based control strategies. To address this, a two-layer safe deep reinforcement learning (DRL) framework is proposed to coordinate PDRC envelopes and HVAC systems for enhanced energy savings while maintaining thermal comfort. In the training layer, a Lagrange multiplier is incorporated into the reward function to dynamically penalize comfort violations, enabling the agent to learn a weather-adaptive trade-off between comfort and energy use. In the deployment layer, a physics-based predictor and optimizer is used to reduce the risk of thermal comfort violations. A comparative study of building energy performance under different control strategies is conducted. Results show that PDRC envelopes significantly reduce the interior surface temperatures of roofs and walls, improving thermal comfort even at the same indoor air temperature. Compared with rule-based control, the proposed framework further reduces cooling-season electricity consumption and cost by 12.1% and 13.5%, respectively, while also outperforming standard DRL control in thermal comfort. This study provides new insight into intelligent control of buildings with PDRC envelopes.