中国电力 ›› 2021, Vol. 54 ›› Issue (2): 44-51.doi: 10.11930/j.issn.1004-9649.202006061

• 国家“十三五”智能电网重大专项专栏:(五)电力传感技术及应用专栏 • 上一篇    

基于残差卷积神经网络的开关柜局部放电模式识别

黄雪莜1, 熊俊1, 张宇1, 刘辉1, 陈鹭1, 孟祥麟2, 江秀臣2   

  1. 1. 广州供电局有限公司荔湾供电局,广东 广州 510150;
    2. 上海交通大学 电气工程系,上海 200240
  • 收稿日期:2020-06-02 修回日期:2020-12-14 发布日期:2021-02-06
  • 作者简介:黄雪莜(1988-),女,硕士,工程师,从事配网运行、带电检测技术、电能质量研究,E-mail:495357689@qq.com;熊俊(1983-),男,硕士,高级工程师,从事GIS在线监测技术、气体放电理论研究,E-mail:1311966654@qq.com;江秀臣(1965-),男,教授,博士生导师,从事高压设备在线监测、状态检修和电气设备自动化研究,E-mail:xcjiang@sjtu.edu.cn
  • 基金资助:
    中国南方电网公司科技项目(配电设备状态检测数据规范化与高效处理关键技术研究与应用,082100KK52190004)

Partial Discharge Pattern Recognition of Switchgear Based on Residual Convolutional Neural Network

HUANG Xueyou1, XIONG Jun1, ZHANG Yu1, LIU Hui1, CHEN Lu1, MENG Xianglin2, JIANG Xiuchen2   

  1. 1. Liwan Power Supply of Guangzhou Power Supply Co. Ltd, Guangzhou 510150, China;
    2. Department of Electrical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China
  • Received:2020-06-02 Revised:2020-12-14 Published:2021-02-06
  • Supported by:
    This work is supported by Science and Technology Project of CSG (Research and Application of Key Techniques for Standardization and Efficient Processing of Distribution Equipment Condition Detection Data, No.082100KK52190004)

摘要: 传统的开关柜局部放电模式识别方法缺乏一定的泛化性能且识别准确率低,难以在实际工程中应用。提出了一种基于残差卷积神经网络的开关柜局部放电模式识别方法,通过在网络中加入残差模块以解决随着网络层数加深导致准确度饱和后出现退化的问题,并综合利用开关柜局部放电数据的浅层与深层特征融合学习,实现模式识别。通过开关柜不同绝缘缺陷类别的局部放电模拟实验与配电站现场检测,构建了开关柜局部放电数据样本库,并进行了实验分析。实验结果表明:所提方法的识别正确率达96.06%,相比传统识别方法至少提高了20.22%,且随着训练集样本数量的增加,识别率有更大提升。综合使用特征层融合模块和残差模块,显著提升了模型的泛化性能,更适用于实际工程。

关键词: 卷积神经网络, 残差模块, 特征层融合, 局部放电, 模式识别

Abstract: The traditional switchgear partial discharge pattern recognition methods lack certain generalization performance and have low recognition accuracy, so it is hard to apply them in practical engineering. A method for switchgear partial discharge pattern recognition based on residual convolutional neural network is proposed. A residual module is added to the network to solve the degradation problem after accuracy saturation caused by the deepening of network layer number, and the shallow and deep feature fusion learning of switchgear partial discharge data is integrated to achieve the pattern recognition. Based on the partial discharge simulation experiments of different insulation defect categories of switchgears and the field testing of distribution stations, a sample database of switchgear partial discharge data is constructed, and experimental analysis is conducted. The experimental results show that the recognition correct rate of the proposed method reaches 96.06%, at least 20.22% higher than that of the traditional recognition methods, and the recognition rate can be improved more with the increase of the number of samples in the training set. Through integrated use of the feature layer fusion module and residual module, the proposed model is significantly improved in the generalization performance and is more suitable for practical engineering.

Key words: convolutional neural network, residual module, feature layer fusion, partial discharge, pattern recognition