考虑机理影响下基于图神经网络的电力系统异常数据检测算法OA

With the expansion of a new type of power system and increased topological complexity,anomaly detection faces challenges such as weak spatiotemporal correlations,imbalanced class distributions,and noise interference.Traditional methods are difficult to effectively integrate the physical mechanisms and topological characteristics of power systems.Therefore,this paper proposes a mechanism-aware GCN-GAT-Hybrid graph neural network for anomaly detection.First,a hybrid module combining graph convolutional networks(GCN)and graph attention networks(GAT)is designed,leveraging residual connections to fuse GCN's structural priors with GAT's dynamic neighborhood weighting,enhancing feature extraction for sparse node relationships and high-order semantics.Second,a topology-constrained graph enhancement method optimizes adjacency matrices using node physical attributes,suppressing noise propagation.Furthermore,a multi-level fusion mechanism integrates global pooling and dynamic classifiers to coordinate node-level anomaly localization and system-level semantic modeling.Experiments on the IEEE14-bus system demonstrate that the proposed algorithm achieves average F1-scores of 96.17%,92.28%,and 86.64%under low,medium,and high noise levels.This study provides an intelligent solution for power system security monitoring that harmonizes physical mechanisms and data-driven approaches.