基于深度-迁移学习的输电线路故障选相模型及其可迁移性研究

为提高输电线路故障诊断模型的可迁移性，根据迁移学习理论将输电线路分为源线路和目标线路，提出一种基于深度-迁移学习的输电线路故障类型识别方法。通过组合不同故障条件，生成输电线路故障期间的时序数据，并通过对数据的预处理，得到面向卷积神经网络的输入数据样本；利用源域数据对初始卷积神经网络进行预训练，获取适用于源线路故障类型识别的预训练模型；采用最大均值差异法对源线路和目标线路进行相似性检验，筛选出待迁移的源域预训练模型；利用目标域数据对预训练模型进行微调迁移训练，获取最终的目标域故障诊断模型。仿真结果表明，利用源域数据量5 %的目标域数据对预训练模型进行微调迁移训练，得到的目标域模型对目标线路故障诊断的准确率达99 %以上。

Transmission line fault phase selection model based on deep-transfer learning and its transferability

In order to improve the transferability of transmission line fault diagnosis models, the transmission lines are divided into source lines and target lines based on transfer learning theory, then a method based on deep-transfer learning for identifying transmission line fault types is proposed. The time series data during transmission line faults is generated by combining different fault conditions, and the input data samples of CNN(Convolutional Neural Network) are obtained by data preprocessing. Then the initial CNN is pre-trained by using the source domain data to obtain a pre-trained model of the source line fault type identification. Next, the maximum mean difference method is used to test the similarity of source and target lines, and the source domain pre-trained model to be migrated is screened out. The target domain data is used to fine-tune the migration training to obtain the final target domain fault diagnosis model. The simulative results show that by using the target domain data of 5 % of the source domain data to fine-tune the migration training of the pre-trained model, the target line fault diagnosis accuracy of the target domain model can reach more than 99 %.