基于PCP和改进CNN的变压器内外故障诊断方法研究  

Research on Internal and External Fault Diagnosis Method of Transformer Based on PCP and Improved CNN

作　　者：何龙 朱咏明 马金财 刘荣 HE Long;ZHU Yongming;MA Jincai;LIU Rong(State Grid Xinjiang Electric Power Co.,Ltd.,Changji Power Supply Company,Xinjiang Changji 831100,China)

机构地区：[1]国网新疆电力有限公司昌吉供电公司,新疆昌吉831100

出　　处：《电力电容器与无功补偿》2025年第2期106-119,共14页Power Capacitor & Reactive Power Compensation

基　　金：国家电网公司科技项目(5230CJ220001)。

摘　　要：为了有效防止变压器差动继电器在非故障瞬态情况下的误动,提出了一种区分变压器内部、外部故障及其暂态过程的快速辨识方法。首先,针对工程现场监测数据可能包含噪声的问题,提出一种利用平行坐标图(PCP)对三相差动时间序列图像化的方法,将三相时间序列可视化,容纳了故障所有信息的同时保障了信号的抗噪性;其次,对卷积神经网络(CNN)加以改进:在特征提取部分增加ECA注意力机制模块,再利用优化后的多核支持向量机对CNN自带Softmax分类器进行改进,增强CNN对各类故障场景分类的适用性,并与实现变压器内外故障的诊断与辨识;最后,搭建变压器内部和外部故障仿真模型,获取绕组匝间故障、接线故障以及励磁涌流、电流互感器(CT)饱和铁磁谐振等故障时差动保护监测数据,据此构建图像化故障样本集,对论文所提方进行验证,并与同类算法进行对比,证明所提方法的优越性。In order to effectively prevent the maloperation of transformer differential relays in non fault transient conditions,a kind of fast identification method to distinguish internal and external faults and their transient processes in transformers is proposed.Firstly,a method to visualize the three-phase differential time series is proposed by using Parallel Coordinate Plot(PCP)to address the issue of potential noise in on-site monitoring data,which,at the same time of containing all fault information,ensures signal noise resistance.Secondly,improvements are made to the CNN:an ECA attention mechanism module is added to the feature extraction section,and the optimized multi kernel support vector machine is used to improve the softmax classifier of the convolutional neural network(CNN),enhancing the applicability of CNN to various fault scenarios classification and facilitating the diagnosis and identification of internal and external faults in transformers.Finally,a internal and external fault simulation model of transformer is set up to obtain such differential protection monitoring data as winding inter turn faults,wiring faults,excitation inrush currents,current transformer(CT)saturation ferromagnetic resonance and other faults.Based on this,a graphical fault sample set is set up,the proposed method in this paper is verified and compared with similar algorithms so to prove the superiority of the proposed method.

关 键 词：变压器故障 平行坐标图 卷积神经网络 

分 类 号：TM41[电气工程—电器]