Enhancing resilience and prioritizing critical infrastructure through deep reinforcement learning-based network reconfiguration in islanded microgrid networks

The increasing reliance on electrical energy necessitates resilient power distribution systems capable of withstanding disruptions caused by natural disasters or cyber-attacks. This paper presents a novel dynamic network reconfiguration framework to enhance the resilience of power distribution systems by forming islanded microgrids during main grid failures. The framework utilises a reinforcement learning approach based on deep Q-networks (DQN) to dynamically optimise the operation of remote-controlled switches for network reconfiguration. A convolutional neural network is integrated into the DQN architecture for Q-value estimation, enabling efficient decision-making under varying operational scenarios. The proposed method is validated using a modified CIGRE medium-voltage distribution network, considering diverse scenarios of load demand and generation capacity. Results demonstrate the framework's capability to prioritise critical loads, maintain system stability, and ensure compliance with operational constraints, such as voltage profiles and line loading limits. The reconfiguration process dynamically forms islanded microgrids, effectively balancing supply and demand while minimising power losses and reducing switching operations. This study underscores the potential of deep reinforcement learning in enhancing grid resilience. Future work aims to incorporate risk-aware learning to address uncertainties, further safeguarding power distribution networks as critical infrastructure.