Medium Voltage Direct Current Shipboard Power Network Reconfiguration Using Graph-Based Reinforcement Learning

Abstract Future marine vessels are set to operate with greater autonomy, reducing the need for onboard crew members. Artificial intelligence offers promising solutions to achieve this, especially for real-time decision-making. This article presents a novel approach for the automatic reconfiguration of shipboard power networks using graph-based reinforcement learning (RL). We modeled the medium voltage direct current (MVDC) shipboard power network reconfiguration problem as a markov decision process, utilizing deep Q-learning to determine the optimal network control policy. The RL policy network is designed using a graph convolutional network (GCN). This technique optimizes the optimal status (ON/OFF) of switches in the MVDC shipboard power network, ensuring maximum power availability to loads during disruptive events such as network faults. The developed method is validated on a four-zone MVDC shipboard power network, demonstrating its effectiveness in autonomously and efficiently managing power distribution. By leveraging artificial intelligence (AI)-driven approaches, shipboard power networks can achieve adaptive and intelligent reconfiguration, enhancing resilience, reducing response times, and mitigating the cascading risks associated with manual intervention. Our approach aligns with the broader vision of next-generation autonomous ships, where data-driven real-time decision-making is promoted for operational efficiency and resilience.