Model-Based Safe Reinforcement Learning for Active Distribution Network Scheduling

Data-driven methods, especially reinforcement learning (RL), are adept at addressing uncertainties but are poor at ensuring safety, which is a critical requirement in active distribution networks (DNs). To address the problem of active DN scheduling and to overcome RL’ most critical drawback—security risk, this paper proposes a model-based safe RL framework that embeds a model-based safety module (MBSM) in the RL’s loop. The proposed framework can guarantee that the agent’s actions (real/reactive power outputs of controllable distributed energy resources (DERs)) strictly satisfy the DN’s operational security constraints. It does not rely on any expert knowledge and is suitable for application in large-scale systems. Comparative studies against existing Safe RL (SRL) and classic optimization methods verify that the proposed method achieves the best performance in terms of DERs operating cost and renewable energy consumption while strictly satisfying the DN’s operational security constraints.