基于卷积神经网络的交直流输电系统故障诊断  

Fault Diagnosis for AC/DC Transmission System Based on Convolutional Neural Network

作　　者：张大海[1] 张晓炜 孙浩 和敬涵[1] ZHANG Dahai;ZHANG Xiaowei;SUN Hao;HE Jinghan(School of Electrical Engineering,Beijing Jiaotong University,Beijing 100044,China;Changzhi Power Supply Company of State Grid Shanxi Electric Power Company,Changzhi 046000,China)

机构地区：[1]北京交通大学电气工程学院,北京市100044 [2]国网山西省电力公司长治供电公司,山西省长治市046000

出　　处：《电力系统自动化》2022年第5期132-140,共9页Automation of Electric Power Systems

基　　金：国家重点研发计划资助项目(2016YFB0900600)。

摘　　要：随着交直流输电系统规模的不断扩大,电网结构和故障特征愈加复杂,现有故障诊断方法面对复杂电网和超大数据量时难以精准提取故障特征,急需适应性强且准确率高的电网故障诊断方法。为此提出一种基于卷积神经网络(CNN)的电网故障诊断方法。首先,通过逐层筛选、逐层增叠的网络构造方式逐步测试,其目的是为了构建充分适应于电网故障诊断的网络结构;然后,利用网络层级优化策略调整训练参数,并以交叉熵最小为目标对深层故障特征进行挖掘;最后,在MATLAB/Simulink平台上搭建交直流输电系统模型,结合t分布随机邻域嵌入(t-SNE)可解释性技术展示诊断效果,通过与传统方法对比证明所提方法能够深度挖掘故障特征且具备很高的诊断准确率。With the continuous expansion of the scale of AC/DC transmission systems, the power grid structure and fault characteristics become more and more complex. The existing fault diagnosis methods confront difficulty in accurately extracting fault characteristics in the face of complex power grid and large amount of data. There is an urgent need for power grid fault diagnosis methods with strong adaptability and high accuracy. Therefore, a convolutional neural network(CNN) based power grid fault diagnosis method is proposed. Firstly, it is tested step by step through the network construction mode of layer-by-layer screening and layer-by-layer superposition to build a network structure fully suitable for power grid fault diagnosis. Then, the network level optimization strategy is used to adjust the training parameters, and the deep fault features are mined with the goal of minimizing the cross entropy. Finally, the AC/DC transmission system model is built on MATLAB/Simulink platform, combined with t-distributed stochastic neighbor embedding(t-SNE) interpretability technology to show the diagnosis effect. Compared with traditional methods, it is proven that the proposed method can deeply mine fault characteristics and has high diagnosis accuracy.

关 键 词：深度学习 卷积神经网络 交直流输电系统 故障诊断 t分布随机邻域嵌入 

分 类 号：TM721.3[电气工程—电力系统及自动化] TP183[自动化与计算机技术—控制理论与控制工程]