摘要
主流的配电网故障定位和识别故障类别任务分开建模,却忽略了故障定位和故障类别之间关联性,该文提出了故障定位以及故障类别的端到端联合模型。利用契比雪夫图卷积神经网络(ChebNet)作为编码器,聚合静态的图结构和动态的节点信息得到各节点的数学表达;在故障定位解码端,通过多头自注意力机制建立适用于节点属性变化以及融合配电网拓扑结构的配电网故障定位模型;在故障分类解码端,结合故障定位解码端的故障区域信息以及ChebNet编码器得到的各节点的数学表达,通过全连接层建立故障类型识别模型。实验结果表明,基于契比雪夫图卷积神经网络在双电源配电网中故障定位中效果较好,故障定位准确率达到98.25%,故障类别任务中的准确率为93.11%。该方法适用于主动配电网结构灵活及含分布式电源的配电网络中。
The mainstream distribution network fault location and fault category identification tasks are modeled separately,but the correlation between fault location and fault category is ignored.In this paper,a joint end-to-end model for fault location as well as fault category is proposed.The Chebyshev graph convolutional neural network(ChebNet)is used as an encoder to aggregate the static graph structure and dynamic node information to obtain the mathematical representation of each node.At the fault location decoder,a distribution network fault location model applicable to node attribute changes and fused distribution network topology is established through multi-head self-attention.At the fault classification decoder,combined with the fault area information of the fault location decoder and the mathematical expression of each node of ChebNet encoder,the fault type recognition model is established through fully connected layer.The experimental results show that Chebyshev graph-based convolutional neural network works have better effect in fault location in dual power distribution networks,with the accuracy rate of fault location reaching 98.25%and the accuracy rate of fault classification task 93.11%.The method is applicable in distribution networks with flexible active distribution network structures and distributed power supply.
作者
许可
范馨月
张恒荣
XU Ke;FAN Xinyue;ZHANG Hengrong(Faculty of Mathematics,School of Mathematics and Statistics,Guizhou University,Guiyang 550025,China;Electric Power Research Institute,Guizhou Power Grid Co.,Ltd.,Guiyang 550002,China)
出处
《实验技术与管理》
CAS
北大核心
2023年第1期26-30,共5页
Experimental Technology and Management
基金
贵州省科技计划项目(黔科合平台人才〔2020〕5016)
贵州大学教改项目(XJG2021027)
贵州大学一流课程培育项目(XJG2021040)。
关键词
契比雪夫图卷积神经网络
多头自注意力机制
配电网
联合模型
Chebyshev graph convolutional neural network
multi-headed self-attentive mechanism
distribution network
joint model
