Service Restoration Using Deep Reinforcement Learning and Dynamic Microgrid Formation in Distribution Networks

This is the authors' abstract. We don't add key points for this paper.

A resilient power distribution network can reduce length and impact of power outages, maintain continuous services, and improve reliability. One effective way to enhance the system's resilience is to form microgrids during outages. In this article, a novel dynamic microgrid formation-based service restoration method using deep reinforcement learning is proposed, and it is treated as a Markov decision process (MDP) while taking operational and structural limitations of microgrids into account. The deep Q-network is employed to obtain optimal control strategies for microgrid formation. We have introduced a new way for the agent to choose actions when building a microgrid using the deep Q-learning method, which ensures that the microgrid has a feasible radial structure. The proposed service restoration method enables real-time computing to facilitate online formation of dynamic microgrids and adapts to changing conditions. The influence of optimal switch placement on service restoration using proposed method is also investigated. The effectiveness of proposed service restoration method is validated by case studies using the modified IEEE 33-node test system and a real 404-node distribution system operated by Saskatoon Light and Power in Saskatoon, Canada.