Collaborative multi-agent dispatch with adaptive scenario recognition via entropy-regularized DRL for robust power system operation

To overcome the limitations of centralized, single-agent scheduling models in capturing diverse operational scenarios—which leads to insufficient robustness and low flexibility in real-time decision-making—this paper proposes a multi-agent collaborative optimal dispatch method for modern power systems. First, a data-driven, multi-scenario learning framework integrates multi-agent training and multi-task coordination with scenario recognition to enable adaptive scenario transitions for collaborative optimal dispatch. Next, the paper proposes a dispatch agent model using an entropy-regularized deep reinforcement learning (DRL) algorithm. This model iteratively couples power flow computations with DRL training to learn sub-dispatch agents corresponding to representative operational scenarios. Finally, a multi-agent dispatch coordination algorithm, leveraging multilayer perceptron (MLP)-based scenario recognition, proactively identifies the scenario labels of current operating conditions to rapidly coordinate relevant sub-agents for real-time dispatch tasks. Case studies validate the robustness of the multi-task modeling framework and the superiority of the produced dispatch strategies.