基于时空循环图神经网络的电网风险诊断框架  

A Risk Diagnosis Framework for Power Grids Based onSpatial-Temporal Recurrent Graph Neural Network

作　　者：胡翔 陈胜 何强 吴育全 龚正宇 HU Xiang;CHEN Sheng;HE Qiang;WU Yuquan;GONG Zhengyu(Southwest Branch of State Grid Corporation of China,Chengdu 610041,China)

机构地区：[1]国家电网有限公司西南分部,四川成都610041

出　　处：《无线电工程》2024年第9期2240-2248,共9页Radio Engineering

基　　金：国家电网有限公司西南分部科技项目(SGSW0000DDKZZXJS2200072)。

摘　　要：风险诊断有益于明确风险隔离框架以及指定系统恢复措施,可为电网检修方式安排提供决策参考。在分布式能源集成化程度日益提高的情况下,基于逆变器的发电机异常电流较小,传统的继电保护失效,对电网系统的风险感知提出了新的要求。提出一种基于时空循环图神经网络(Spatial-Temporal Recurrent Graph Neural Network,STRGNN)的电网风险诊断框架,提升了检修方式的风险识别能力。STRGNN可以从关键母线上的电压测量单元数据中提取时空特征,根据特征进行风险事件检测、风险类型/相位分类、风险定位等操作。与现有研究成果相比,STRGNN对风险诊断具有更好的泛化能力。此外,STRGNN提取电压信号而不是电流信号,不需要在电网系统的所有线路上安装继电器,不受电流测量单元数量的约束。在波茨坦微电网系统和IEEE-123节点馈线系统上进行大量实验,结果表明STRGNN相比其他基准方法具有更好的性能。相较于最先进的图卷积方法在IEEE-123节点馈线系统上,风险定位准确率提升了1.8%。Risk diagnostics are useful for clarifying risk isolation scenarios and specifying system recovery actions,which can provide decision-making reference for the arrangement in the maintenance mode of power grids.With the increasing of distributed energy integration,the inverter-based generator has less abnormal current and the traditional relay protection fails,which raises new requirements for the risk perception of power grid systems.A risk diagnosis framework based on Spatial-Temporal Recurrent Graph Neural Network(STRGNN)for power grids is proposed,which improves the risk identification ability of maintenance methods.STRGNN can extract temporal and spatial features from the voltage measurement unit data on the key bus.Based on these features,risk event detection,risk type/phase classification,risk localization,and other operations are performed.Compared with previous research results,STRGNN has better generalization ability for risk diagnosis.In addition,STRGNN extracts voltage signals rather than current signals,so there is no need to install relays on all lines of the grid system and STRGNN is not constrained by the number of current measuring units.Extensive experiments on the Potsdam Microgrid System and IEEE-123 node feeder system show that STRGNN has better performance than other benchmark methods.When compared with the most advanced graph convolution method,the accuracy of risk location has increased by 1.8%on the IEEE-123 node feeder systems.

关 键 词：风险诊断 时空循环图神经网络 时空特征 电压信号 微电网 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]