Safe Reinforcement Learning for Joint Lighting and Blind Control Using Predictive Control Barrier Functions

Lighting and shading systems bear enormous potential for reducing building energy consumption, yet their coordinated control remains challenging due to competing objectives between energy efficiency and occupant comfort. While Reinforcement Learning (RL) offers a promising approach, standard algorithms lack formal safety guarantees, leading to potential comfort violations during both training exploration and real-world deployment. This paper proposes a safe RL framework for joint control of continuous electric lighting and discrete automated blinds using Predictive Control Barrier Functions (PCBF). We formulate a two-stage Mixed-Integer Linear Program (MILP) safety filter that explicitly handles hybrid action spaces while enforcing illuminance and glare constraints. The reward function evaluates both the agent’s proposed action and its deviation from the safe correction to allow the policy to learn constraint boundaries while executing safety-filtered actions. We integrate this framework with a physics-based simulation environment utilizing Radiance’s three-phase method to capture directional light distribution for efficient calculation of Simplified Daylight Glare Probability (DGPs). Comparative evaluation against safety-focused MPC, RL, and rule-based controllers demonstrates that the proposed framework achieves high constraint satisfaction while consuming significantly less energy than safety-focused MPC, offering a balanced trade-off between safety and energy efficiency.