摘要
电网拓扑结构复杂、分支众多、潮流分布不平衡、故障样本较少且难以获取。为提高配电网的故障诊断准确性,提出将迁移学习的思想与卷积神经网络(convolutional neural networks,CNN)相结合,以此来解决目标域样本不足导致训练效果差的问题,同时利用主成分分析(principal component analysis,PCA)对时序数据进行降维,提升运行速率,形成配电网故障诊断方法。首先对PCA和CNN的结构特点进行分析;然后通过仿真模拟不同的故障条件,生成面向CNN的时序数据。再通过最大均值差异法(MMD)选择出最适合迁移的源域数据,建立源域故障识别的预训练模型。最后使用目标域数据,在预训练模型的基础上进行迁移微调训练,得到故障诊断模型。仿真结果表明,该方法能够在小样本的情况下迅速完成对故障类型的精准预测。
The power grid topology is complex,with many branches and unbalanced power flow distribution,which makes the fault samples less and difficult to obtain.To improve the fault diagnosis accuracy of distribution network,the combination of transfer learning and convolutional neural networks(CNN)was proposed to solve the problem of poor training effect caused by insufficient samples in the target domain.Meanwhile,Principal component analysis(PCA)was used to reduce the dimension of time series data and to improve the operation speed to form a fault diagnosis method for distribution network.Firstly,the structural characteristics of PCA and CNN were analyzed,and different fault conditions were simulated to generate time series data for CNN.Secondly,the most suitable source domain data for migration was selected by the maximum mean difference method(MMD),and a pre-training model for source domain fault identification was established.Finally,on the basis of the pre-trained model,the target domain data was migrated fine-tuning training to obtain a fault diagnosis model.The simulation results show that the method can complete the accurate prediction of fault types quickly in the case of small samples.
作者
丁津津
邵庆祝
齐振兴
谢民
高博
于洋
DING Jin-jin;SHAO Qing-zhu;QI Zhen-xing;XIE Min;GAO Bo;YU Yang(State Grid Anhui Electric Power Research Institute, Hefei 230601, China;State Grid Anhui Electric Power Co., Ltd., Hefei 230022, China;School of Electrical Engineering and Automation, Anhui University, Hefei 230601 , China)
出处
《科学技术与工程》
北大核心
2022年第14期5653-5658,共6页
Science Technology and Engineering
基金
国家自然科学基金(52077001)。
关键词
配电网
故障识别
迁移学习
卷积神经网络
主成分分析
distribution network
fault identification
transfer learning
convolutional neural networks
principal component analysis
