Achieving carbon neutrality by 2050 will require power-grid operators to absorb unprecedented volumes of variable solar and wind generation while maintaining reliability.To tackle this systems-level bottleneck,Ré...Achieving carbon neutrality by 2050 will require power-grid operators to absorb unprecedented volumes of variable solar and wind generation while maintaining reliability.To tackle this systems-level bottleneck,Réseau de Transport d’Électricité(RTE)and the research community launched Learn To Run A Power Network(L2RPN),a crowd-sourced competition aiming to accelerate the integration of intermittent renewables into power-grid operations.L2RPN is based on 16 years of weekly scenarios(832 in total)on a 118-node grid under realistic constraints,and casts real-time grid operation as a Markov-Decision-Process.The six participating teams tackled the challenge by developing autonomous agents with various strategies blending heuristics,optimization,data scaling,supervised learning,and reinforcement learning.We provide a detailed overview of all six participants’performance under the competition’s demanding design.In addition,we present an in-depth analysis of the winning solution-made publicly available-which achieves consistent decision making across scenarios,executes real-time multimodal actions in under five seconds,and performs efficient topology control via action-space reduction and a neural policy that predicts useful grid actions with over 80%accuracy.In parallel,we trained a neural alert module on 315,000 samples derived from top agents,achieving 93.9%recall in flagging dangerous states and allowing agents to predict future failure.Finally,this work not only demonstrates AI’s promise and current limits in real-time grid management but also lays a transparent foundation for more robust,trustworthy systems in the energy transition.展开更多
文摘Achieving carbon neutrality by 2050 will require power-grid operators to absorb unprecedented volumes of variable solar and wind generation while maintaining reliability.To tackle this systems-level bottleneck,Réseau de Transport d’Électricité(RTE)and the research community launched Learn To Run A Power Network(L2RPN),a crowd-sourced competition aiming to accelerate the integration of intermittent renewables into power-grid operations.L2RPN is based on 16 years of weekly scenarios(832 in total)on a 118-node grid under realistic constraints,and casts real-time grid operation as a Markov-Decision-Process.The six participating teams tackled the challenge by developing autonomous agents with various strategies blending heuristics,optimization,data scaling,supervised learning,and reinforcement learning.We provide a detailed overview of all six participants’performance under the competition’s demanding design.In addition,we present an in-depth analysis of the winning solution-made publicly available-which achieves consistent decision making across scenarios,executes real-time multimodal actions in under five seconds,and performs efficient topology control via action-space reduction and a neural policy that predicts useful grid actions with over 80%accuracy.In parallel,we trained a neural alert module on 315,000 samples derived from top agents,achieving 93.9%recall in flagging dangerous states and allowing agents to predict future failure.Finally,this work not only demonstrates AI’s promise and current limits in real-time grid management but also lays a transparent foundation for more robust,trustworthy systems in the energy transition.
