A multi-rate sampling PMU-based event classification in active distribution grids with spectral graph neural network
A multi-rate sampling PMU-based event classification in active distribution grids with spectral graph neural network
复制标题
DOI：
10.1016/j.epsr.2022.108145
复制
发表时间：
2022
影响因子：
3.9
通讯作者：
Mohammad MansourLakouraj;Mukesh Gautam;H. Livani;M. Benidris
中科院分区：
文献类型：
--
作者：
Mohammad MansourLakouraj;Mukesh Gautam;H. Livani;M. Benidris
相量测量单元 (PMU) 是过去二十年在传输网络中激增的时间同步测量设备。最近，人们努力将该技术引入配电网，用于不同的应用，例如三相状态估计、故障和事件分析以及相位识别。流式传输的时间同步电压和电流相量数据可用于沿配电馈线进行事件分类和区域识别，以确定事件的类型和位置，这是任何故障和事件检测、定位和隔离软件的重要功能。本文采用基于谱理论的图卷积进行事件分类和区域识别。所提出的模型使用改进的图卷积滤波器来聚合 PMU 数据的区域多速率样本，即来自多个节点的电压幅度和角度。除了测量节点的这些时间数据之外，包含边缘特征的网络的物理配置也被赋予图卷积网络（GCN），不仅可以对事件类型进行分类，还可以识别受影响的区域和位置。所提出的基于图形的方法在标准测试系统上进行了测试，其中包括电容器和分布式能源相关事件、稳压器故障、负载突然变化以及不同类型的故障。使用准确度、召回率、精度和 F-1 得分指标将结果与基线方法、切比雪夫图神经网络 (GNN)、决策树、逻辑回归和 K 最近邻进行比较。此外，还针对安装的 PMU 数量、测量噪声水平、可用历史数据的大小、网络边缘功能的可用性以及 GNN 的不同设计进行了性能敏感性分析。
Phasor measurement units (PMUs) are time-synchronized measurement devices that have been proliferated in transmission networks during the last two decades. Recently, there have been efforts to bring this technology to distribution grids for different applications such as three-phase state estimation, fault and event analyses, and phase identification. Streamed time-synchronized voltage and current phasor data can be used for events classification and region identification along distribution feeders to determine the type and location of events, which are important features of any fault and event detection, location, and isolation software. In this paper, the spectral theory-based graph convolution is used for event classification and region identification. The proposed model uses modified graph convolution filters to aggregate the regional multi-rate samples of PMU data, i.e., voltage magnitude and angles from several nodes. Besides these temporal data of the measured nodes, the physical configuration of the network containing edge features are given to the graph convolution network (GCN) to not only classify the event type, but also identify the affected region and location. The proposed graph-based method is tested on a standard test system with capacitor and distributed energy resources-related events, malfunction of voltage regulator, sudden load changes, and different types of faults. The results are compared with baseline methods, Chebyshev graph neural network (GNN), decision tree, logistic regression and K-nearest neighbor using the accuracy, recall, precision and F-1 score metrics. Furthermore, performance sensitivity analysis is carried out with respect to the number of installed PMUs, measurement noise level, size of available historical data, availability of network edge features, and different designs of GNN.