Safe deep reinforcement learning for building energy management

The optimization of building energy systems poses a complex challenge due to the dynamic nature of building environments and the need for ensuring both energy efficiency and system reliability. This paper proposes a novel approach that synergizes deep reinforcement learning (DRL) with model predictive control (MPC) for safe and efficient building energy management. The approach, named safe DRL, leverages the DRL to explore optimality in uncertain environments and the MPC to dynamically ensure constraints. Off-the-shelf DRL tools are utilized in the safe DRL approach which avoids case-by-case and effort-intensive hyperparameter tuning. MPC with only an approximate building model is then adopted to efficiently filter the DRL action candidates to ensure the safety of the constraints. Additionally, a comprehensive safety metric is proposed considering the duration and severity of constraint violation. Extensive case studies with high-fidelity building models show that the safe DRL approach significantly enhances the constraint satisfaction and improves building performance at the cost of slightly more decision time of a few seconds.