Dynamic Power Flow Analysis and Fault Characteristics: A Graph Attention Neural Network
To address the challenges of scarce labeled data, emergence of unknown data due to newly integrated devices, and the trade-off between computational efficiency and diagnostic accuracy in dynamic power flow analysis and fault diagnosis for smart grids, this paper proposes a semi-supervised graph learning framework. The method innovatively integrates adaptive neighborhood clustering (CAN) with a Markov random field (MRF) to construct a physics-informed graph structure, and employs a graph attention network (GAT) to model label dependencies among nodes—thereby reducing reliance on large-scale labeled datasets. Evaluated on real-world grid scenarios, the framework achieves significant improvements in computational efficiency and, under few-shot settings, yields an average 12.6% gain in F1-score. It thus enables high-accuracy, low-label-cost, and robust fault diagnosis with strong generalization capability.