Simulation study demonstrates improved operating costs and fewer voltage violations in active distribution networks, highlighting the value of adaptive graph reinforcement learning.
Key Points
Develop a state-adaptive, topology-aware continuous power dispatch framework using multi-head graph attention and deep deterministic policy gradients to mitigate renewable energy and load uncertainty in distribution grids.
Embedded a multi-head graph attention network into a centralized Actor–Critic training architecture (GAT-DDPG) to dynamically adjust spatial message-passing weights according to operational states.
Evaluated performance using extensive continuous-dispatch simulations on a modified IEEE 33-bus active distribution network across a 125-day unseen test set.
Achieved lower comprehensive operating costs and reduced voltage violations compared with standard deep reinforcement learning baseline algorithms.
Tracked dynamically shifting network vulnerabilities through state-dependent attention shifts, validating dispatch stability during temporary reductions in renewable power generation.