摘要
提出了一种基于图神经网络(GNN)的电网故障检测方法,辅助解决网络运维领域的故障问题.通过构建电网特征抽取模型以及知识图谱,将所提出图神经网络应用于电网的故障节点定位;通过引入时间维度信息,将前后时刻的节点状态辅助当前时刻的故障检测;将图神经网络层的各个节点输出特征进行相关性分析,验证图神经网路对节点特征抽取的有效性.结果表明:所提出的图神经网络能够有效地定位电网故障节点,在仿真故障场景上获得了99.53%的定位准确率,图神经网路对节点的特征建模可以定性观测节点故障扰动传播情况,对故障节点分析起到了重要辅助作用.
A power grid fault detection method based on graph neural network(GNN)was proposed for intelligent fault diagnosis,which can help solve fault problems in the field of network operation and maintenance.The proposed GNN-based neural network was utilized to locate faulty nodes of the power grid by constructing an electrical feature extraction model and a knowledge graph.By introducing the information in the temporal dimension,the node statuses of the previous and next time instances were used to assist the fault detection for the current time instance.The correlation analysis was conducted on the output features of each node in the graph neural network layer to verify the effectiveness of the graph neural network for node feature extraction.The experimental results demonstrate that the proposed method can achieve an accuracy of 99.53%in locating the fault nodes on simulated fault scenarios.From the feature modeling of the graph neural network,the disturbance propagation of the fault nodes can be qualitatively inspected,which plays an important role to facilitate the fault node analysis.
作者
张晓华
袁肖赟
杨红英
杨延栋
赛峰
ZHANG Xiaohua;YUAN Xiaoyun;YANG Hongying;YANG Yandong;SAI Feng(State Grid Jibei Electric Power Co.Ltd.,Beijing 100053,China;Department of Electronic Engineering,Tsinghua University,Beijing 100876,China;China Electric Power Research Institute,Beijing 100192,China;State Grid Xinjiang Electric Power Research Institute,Urumqi 830000,China)
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2023年第3期47-51,59,共6页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家电网有限公司总部科技项目(5100-202155016A-0-0-00)。
关键词
电网故障检测
图神经网络
智能故障诊断
故障定位
知识图谱
fault detection of power grid
graph neural network
intelligent fault diagnosis
fault location
knowledge graph
