基于卷积神经网络的光伏系统直流串联电弧故障检测  

DC series arc-fault detection of photovoltaic system based on convolutional neural network

作　　者：焦治杰 李腾[1] 王莉娜[2] 牟龙华[3] Alexandra Khalyasmaa JIAO Zhi-jie;LI Teng;WANG Li-na;MU Long-hua;Alexandra Khalyasmaa(School of Electrical Engineering, Beijing Jiaotong University, Beijing 100044, China;School of Automation Science and Electrical Engineering, Beihang University, Beijing 100191, China;School of Electronics and Information Engineering, Tongji University, Shanghai 201804, China;Department of Automated Electric Systems, Ural Federal University, Ekaterinburg 620002, Russia)

机构地区：[1]北京交通大学电气工程学院,北京100044 [2]北京航空航天大学自动化科学与电气工程学院,北京100191 [3]同济大学电子与信息工程学院,上海201804 [4]乌拉尔联邦大学电气系统自动化学院,叶卡捷琳堡620002,俄罗斯

出　　处：《电工电能新技术》2019年第7期29-34,共6页Advanced Technology of Electrical Engineering and Energy

基　　金：国家重点研发计划项目(2018YFB1500802)

摘　　要：本文提出一种新颖的基于卷积神经网络的光伏系统直流串联电弧故障检测方法。首先采用短时傅里叶变换提取电流信号的时频信息,以能量谱密度作为电流的时频联合能量函数,构造电流的时频谱图,然后以时频谱图中各时频点的能量谱密度作为卷积神经网络的输入,设计卷积神经网络算法实现电弧故障检测。经实验验证,所提出方法可清晰区分电弧故障电流特征和正常工作电流特征;在实验室测试中,所提出方法可准确地检测出光伏系统直流串联电弧故障。A novel DC series arc-fault detection method for photovoltaic systems based on convolutional neural net- work is proposed. Firstly, the short-time Fourier transform is used to derive the time-frequency information of the current. Then energy spectral density, acted as the time-frequency joint function, is used to construct the time-fre- quency spectrogram of the current. The coordinate information on the time-frequency spectrogram of the current is inputted to the convolutional neural network. And the convolutional neural network is trained to discriminate the arc-fault current and normal operation current. Experiments verified that the proposed method can make a clear dis- tinction between the arc-fault current and normal operation current. In the experiments, the DC series arc-fault cur- rent of the photovoltaic system can be detected accurately.

关 键 词：光伏系统 串联电弧 直流电弧 短时傅里叶变换 卷积神经网络 

分 类 号：TM930.12[电气工程—电力电子与电力传动]