This paper presents a graph reinforcement learning model for outage management in distribution networks to enhance resilience. The model explicitly accounts for network topology variations with switching control and interdependencies between state variables. A Capsule-based graph neural network learns the optimal control policy for power restoration, achieving near-optimal, real-time performance on three test networks (13, 34, and 123-bus modified IEEE networks). Resilience improvement in terms of loss of energy is significant, and the model demonstrates generalizability across various outage scenarios.

Roshni Anna Jacob, Steve Paul, Souma Chowdhury, Yulia R. Gel, Jie Zhang