Multi-Agent Reinforcement Learning via Mean Field Game for Power Grid Fault Diagnosis & Prediction

Abstract

Power grid fault diagnosis and prediction represent critical challenges in modern electrical infrastructure management, requiring rapid, accurate, and scalable decision-making under complex, dynamic conditions. This paper proposes a novel framework integrating MARL with MFG theory to address the inherent scalability limitations of conventional multi-agent approaches in large-scale power systems. Each monitoring agent observes local grid states and interacts with the collective behavior of neighboring agents through a mean field approximation, substantially reducing the computational complexity from exponential to linear scaling with respect to the number of agents. The fault diagnosis and prediction problem are formulated as a stochastic differential game and derive the corresponding mean field equilibrium conditions. Extensive experiments conducted on the IEEE 14-bus, 39-bus, and 118-bus benchmark systems demonstrate that the MFG-MARL framework achieves a fault detection accuracy of 97.3%, 96.1%, and 94.8%, a false positive rate of 1.8%, 2.1%, and 2.7%, separately and prediction lead time of 450 milliseconds. Relative to the strongest MARL baseline evaluated (MAPPO), MFG-MARL improves diagnosis accuracy by 2.3-6.1 percentage points and lowers the false-positive rate by 0.6-2.2 percentage points across the three benchmark systems, while reducing inference latency by 65% on the IEEE 118-bus system (23.4 ms vs. 67.3 ms) , consistent with the linear-scaling argument underlying the mean field approximation.