摘要
为有效诊断高压输变电设备的绝缘状态,保证高压输电系统的安全稳定运行,提出了一种基于多通道卷积神经网络的油纸绝缘局部放电模式识别方法。首先,搭建了工频至低频局部放电试验平台,获取了50、30、5 Hz外施电压下4类典型缺陷的局部放电相位分布图谱,并通过重采样、调幅移相、噪声模拟等对其进行了样本扩充预处理。然后,搭建了具备3通道输入的卷积神经网络,通过该卷积神经网络将不同频率外施电压下同一缺陷的局部放电相位分布图谱进行了通道级无损融合,并对融合后的局部放电相位分布图谱进行了自主特征提取与模式识别。最后,与传统方法的识别结果进行了对比。结果表明:该方法可有效识别油纸绝缘局部放电的类型;相较于单频率单通道卷积神经网络及基于人工特征提取的支持向量机和反向传播神经网络,其对油纸绝缘不同缺陷的局部放电相位分布图谱信息提取得更全面,识别准确率分别提升1.5%、13%、30%;与Vgg、ResNet、DenseNet等深度卷积神经网络相比,该方法的过拟合得到改善,识别准确率分别提高0.25%、1.25%、2.25%,且模型参数文件可节约至少380倍存储空间,便于在边缘计算设备中进行部署。
To effectively diagnose the insulation status of high-voltage power transmission and transformation equipment and ensure the safe and stable operation of the high-voltage power transmission system,a novel method for pattern recognition of partial discharge in oil-paper insulation based on multi-channel convolutional neural network(CNN)was proposed.Firstly,an experiment platform of partial discharge under power frequency to low frequency was built,and the phase resolved partial discharge(PRPD)patterns of 4 types of typical defects under applied voltages at frequency of 50,30 and 5 Hz were acquired,which were then preprocessed by resampling,amplitude and phase transforming,and noise adding.Then,a CNN with 3 input channels was set up,through which the PRPD patterns from the same defect but under applied voltages at different frequencies were combined by channel,and automatic feature extraction and pattern recognition of the combined PRPD patterns were conducted.Finally,the recognition result was compared with that from conditional methods.The result indicates that this method can be adopted to effectively identify the type of the partial discharge in oil-paper insulation.Compared with the single-frequency single-channel CNN and the support vector machine(SVM)and back propagation neural network(BPNN)based on artificial feature extraction,the multi-channel CNN based on automatic feature extraction performs better in recognizing the PRPD patterns of different defects in oil-paper insulation,with higher accuracies of 1.5%,13%and 30%,respectively.The multi-channel CNN performs better than deep CNNs such as Vgg,ResNet and DenseNet due to the over-fitting improvement,with higher accuracies of 0.25%,1.25%and 2.25%,respectively.Additionally,the parameter file of this model is at least 380 times smaller than that of the deep CNNs,which can be easily deployed in edge computing devices.
作者
陈健宁
周远翔
白正
赵云舟
张云霄
张灵
CHEN Jianning;ZHOU Yuanxiang;BAI Zheng;ZHAO Yunzhou;ZHANG Yunxiao;ZHANG Ling(State Key Laboratory of Control and Simulation of Power System and Generation Equipment,Department of Electrical Engineering,Tsinghua University,Beijing 100084,China;The Wind Solar Storage Division of State Key Laboratory of Control and Simulation of Power System and Generation Equipment,School of Electrical Engineering,Xinjiang University,Urumqi 830047,China)
出处
《高电压技术》
EI
CAS
CSCD
北大核心
2022年第5期1705-1715,共11页
High Voltage Engineering
基金
国家重点研发计划(2017YFB0902704)
国家电网有限公司科技项目(5204XQ19004C00K)。
关键词
多通道
卷积神经网络
油纸绝缘
局部放电
模式识别
multi-channel
convolutional neural network
oil-paper insulation
partial discharge
pattern recognition
