电气工程学报 ›› 2020, Vol. 15 ›› Issue (3): 22-30.doi: 10.11985/2020.03.003

• • 上一篇    下一篇

扫码分享

基于卷积神经网络的配电网单相接地故障识别*

李宇(),杨柳林()   

  1. 广西大学电气工程学院 南宁 530004
  • 收稿日期:2020-05-15 修回日期:2020-07-29 出版日期:2020-09-25 发布日期:2020-10-28
  • 作者简介:李宇,男,1995年生,硕士研究生。主要研究方向为配电网自动化技术,深度学习。E-mail:ly_hh_ly@163.com|杨柳林,男,1977年生,博士,副教授,硕士研究生导师。主要研究方向为电力系统分析与计算、深度学习、大数据及其应用等。E-mail:1564652658@qq.com
  • 基金资助:
    * 国家自然科学基金(51667004)

Identification of Single-phase-to-earth Fault in Distribution Network Based on Convolutional Neural Network

LI Yu(),YANG Liulin()   

  1. School of Electrical Engineering, Guangxi University, Nanning 530004
  • Received:2020-05-15 Revised:2020-07-29 Online:2020-09-25 Published:2020-10-28

摘要:

配电网接地故障类型复杂,故障量常呈现间歇性和微弱性的特点,基于传统机器学习方法难以挖掘故障量的有效信息,限制了故障识别性能的提高。针对该问题,提出了一种基于卷积神经网络(Convolutional neural networks, CNN)自动提取故障特征的配电网接地故障识别方法。运用PSCAD软件构建配电网接地故障仿真模型,并由脚本程序生成CNN所需的海量样本集。搭建CNN框架,采用卷积和池化运算提取故障量的特征,以准确地刻画被识别的故障类型,并通过t-SNE方法可视化展示CNN提取故障特征的能力。进行CNN关键参数的调整以及优化算法的优选,以提升模型收敛速度和识别性能。结果表明,与传统机器学习方法相比,CNN模型识别准确率最高,且不受频率、故障位置、电压波动等干扰,并在配电网信息丢失以及注入谐波的情况下,验证了其容错性和适应性。

关键词: 配电网, 故障识别, 卷积神经网络, PSCAD, 数据可视化

Abstract:

The type of ground faults in distribution networks is complex, and the amounts of faults is often intermittent and weak. Based on traditional machine learning methods, it is difficult to mine the effective information of the amounts of faults, which limits the improvement of fault identification performance. Aiming at this problem, a method to identify ground faults based on convolutional neural network (CNN) to extract fault features automatically is proposed. PSCAD software builds a simulation model of the grounding fault of the distribution network, and the massive sample set required by CNN is generated by the script program. Build a CNN framework, use convolution and pooling operations to extract the characteristics of the fault amount to accurately characterize the identified fault type. The ability of CNN to extract fault features is demonstrated through t-SNE method. The adjustment of CNN key parameters and the selection of optimization algorithms are done to improve the model convergence speed and identification performance. The results show that, compared with traditional machine learning methods, the CNN model has the highest identification accuracy and is not disturbed by frequency, fault location, voltage fluctuations, etc. Its fault tolerance and adaptability are verified in the case of the distribution network information loss and harmonic injection.

Key words: Distribution network, fault identification, convolutional neural network, PSCAD, data visualization

中图分类号: