Coordinated dispatch of hybrid mobile power sources for distribution network restoration: A dynamic gradient masking embedded multi-agent meta-deep reinforcement learning method

The co-dispatch of electric vehicles (EVs) and mobile energy storage systems (MESSs) as mobile power sources (MPSs) has emerged as a critical means for the rapid restoration of post-disaster distribution networks (DNs). However, MESSs have DN restoration as a single objective, while private EVs pursue dual objectives: supporting DN restoration and fulfilling individual charging demands. The heterogeneous objectives of EVs and MESSs, coupled with their limited adaptability for rapid deployment in unpredictable post-disaster scenarios, result in coordination failures during deployment. To this end, this paper proposes a dynamic gradient masking embedded multi-agent meta-deep reinforcement learning (DGME-MAMDRL) strategy. A tailored reward function is designed to guide EVs in making correct decisions between charging and DN restoration. Each MPS is modelled as an independent agent within a Markov Game for DN restoration. The proposed strategy is applied to batches of pre-training tasks for knowledge extraction. A dynamic gradient masking mechanism is proposed and embedded within the strategy to enhance the prior knowledge extraction from different tasks. In this way, the agents need only a quick fine-tuning stage for post-disaster deployment. Case studies validate the effectiveness of the proposed strategy in the co-dispatch of hybrid MPSs and its capability for rapid deployment.