High penetrations of distributed energy resources and fast‑varying electric‑vehicle charging cause frequent voltage and line‑loading violations in distribution networks, making safe and low‑cost operation increasingly challenging. To address this, we propose a physics‑informed safe deep reinforcement learning framework that coordinates community battery energy storage to reduce constraint violations while minimizing operating cost.
A dual‑level physics‑informed graph is constructed. The upper level shapes spatial information flow using line impedance and capacity, while the lower level regularizes latent states with power‑balance, voltage‑profile, and flow‑reconstruction losses, enforcing physical consistency relevant to control. This representation captures non‑local spatiotemporal interactions and fully exploits network topology and physical laws.
Based on this graph, we develop Lagrangian Dedicated‑Critic PPO (LDCPPO). LDCPPO employs separate reward and cost critics to achieve constraint‑specific value decomposition; the two critics are merged only during the Lagrangian policy update, preserving heterogeneous learning signals for voltage and line‑loading constraints while applying soft penalty via Lagrange multipliers.
Case studies on networks with high rooftop PV penetration and heavy EV charging demonstrate that LDCPPO reduces the key voltage‑violation magnitude to 19.5% of baseline values, achieves the lowest violation count and line‑overload degree, and maintains competitive operating cost.
Review