基于迁移学习的配电网内部过电压识别方法  

An Internal Overvoltage Identification Method for Distribution Network Based on Transfer Learning

作　　者：徐浩 刘利强[1,2] 吕超 XU Hao;LIU Liqiang;LV Chao(College of Electric Power,Inner Mongolia University of Technology,Hohhot 010080,China;Inner Mongolia Key Laboratory of Electrical Energy Conversion Transmission and Control,Hohhot 010080,China;Inner Mongolia Electric Power Research Institute,Hohhot 010020,China)

机构地区：[1]内蒙古工业大学电力学院,内蒙古呼和浩特010080 [2]内蒙古自治区电能变换传输与控制重点实验室,内蒙古呼和浩特010080 [3]内蒙古电力科学研究院,内蒙古呼和浩特010020

出　　处：《中国电力》2021年第8期52-59,共8页Electric Power

基　　金：内蒙古自治区自然科学基金面上项目(2020MS05029)。

摘　　要：数据驱动方式作为解决配电网内部过电压识别的一种方法,因过电压样本数量较少而在实际应用中受到限制。为此,提出了一种基于迁移学习的深度卷积神经网络(D-CNN)配电网内部过电压识别算法。首先,通过仿真和连续小波变换(CWT)的方法构造了6种10 kV配电网内部过电压二维时频图。然后,分别利用Alexnet、Vgg-16、Googlenet、Resnet50等4种网络模型搭建了基于迁移学习的D-CNN网络模型。最后,将二维时频图带入改造后的D-CNN训练。经对实验结果比较分析发现,新搭建的VGG-16网络识别准确率最高且达到了99.07%,实现了在数据稀缺的情况下过电压故障的准确分类。As a measure for internal overvoltage identification of distribution network,the data driving method is limited in practical applications due to the small number of overvoltage samples.A transfer-learning-based deep convolutional neural network(D-CNN)algorithm is thus proposed to identify the internal overvoltage of distribution network.Firstly,6 types of two-dimension timefrequency maps of 10 kV distribution network internal overvoltage are constructed by simulation and continuous wavelet transform(CWT).Then,the transfer-learning-based D-CNN network models are built using four network models,including Alexnet,Vgg-16,Googlenet and Resnet50.Finally,the two-dimension time-frequency maps are introduced into the transformed D-CNN for training.By comparing and analyzing the experimental results,it is found that the newly constructed VGG-16 network model has the highest identification accuracy,reaching 99.07%,which realizes the accurate classification of overvoltage faults in the case of scarce data.

关 键 词：配电网内部过电压 连续小波变换 迁移学习 深度卷积神经网络 模式识别 

分 类 号：TM73[电气工程—电力系统及自动化]