Multi-Agent Reinforcement Learning for Mobile Energy Resources Scheduling Amidst Typhoons

Substantial threats are posed by typhoon events to critical electrical power infrastructure that can result in human casualties and significant economic losses. Both traditional and renewable power generation systems can be negatively impacted, leading to widespread power outages that compromise public safety. In response, we introduce a novel spatial-topological multi-agent reinforcement learning (ST-MARL) method to optimize post-typhoon power system recovery by leveraging mobile energy resources (MERs). Compared to existing MARL methods, the proposed ST-MARL method utilizes convolutional neural network (CNN) and graph convolutional network (GCN) models to extract spatial-topological information from environmental data like typhoon meteorological conditions and power system states, which facilitates real-time decision-making for the routing and scheduling of MERs. Furthermore, this ST-MARL method achieves a two-stage decision-making process by allocating MERs' rewards with an Alternating Current Optimal Power Flow(AC-OPF) model. This ensures safety and feasibility of the decisions. Additionally, we employ the centralized training and distributed execution (CTDE) paradigm to address coordination challenges among MERs. These approaches collectively aim to enhance coordination among MERs, improve economic efficiency, and ensure the power supply of critical loads during post-typhoon power system recovery. Finally, in our case study of the Hong Kong power network, the results indicate that our ST-MARL method outperforms two main existing MARL methods, achieving an improvement of 5.13% and 6.77% on the reward, respectively.