林济铿 ¹ ，任怡睿 ¹ ，闪 鑫 ^{2, 3, 4}，李 俊 ^{2, 3}，翟明玉 ^{2, 3, 4}，王 波 ^{2, 3}
AuthorsHTML:林济铿 ¹ ，任怡睿 ¹ ，闪 鑫 ^{2, 3, 4}，李 俊 ^{2, 3}，翟明玉 ^{2, 3, 4}，王 波 ^{2, 3}
AuthorsListE:Lin Jikeng¹，Ren Yirui¹，Shan Xin^{2, 3, 4}，Li Jun^{2, 3}，Zhai Mingyu^{2, 3, 4}，Wang Bo^{2, 3}
AuthorsHTMLE:Lin Jikeng¹，Ren Yirui¹，Shan Xin^{2, 3, 4}，Li Jun^{2, 3}，Zhai Mingyu^{2, 3, 4}，Wang Bo^{2, 3}
Unit:1. 同济大学电子与信息工程学院，上海 201804；
2. 南瑞集团(国网电力科学研究院)有限公司，南京 211106；
3. 国电南瑞科技股份有限公司，南京 211106；
4. 智能电网保护和运行控制国家重点实验室，南京 211106
Unit_EngLish:1. College of Electronics and Information Engineering，Tongji University，Shanghai 201804，China；
2. NARI Group Corporation (State Grid Electric Power Research Institute)，Nanjing 211106，China；
3. NARI Technology Co. Ltd.，Nanjing 211106，China；
4. State Key Laboratory of Smart Grid Protection and Control，Nanjing 211106，China
Abstract_Chinese:故障诊断软件作为 SCADA 系统中的标准模块，仍存在着误报率高的问题．针对该问题，本文提出了基于Logistic 回归深度学习的故障诊断新模型及算法，以进一步提高诊断的准确率．该新模型及算法的构建过程如下： 首先，对于每一元件均建立回归型深度学习神经网络(DNN)，其输入为处理成 1/-1 之后的元件故障特征向量，其输出为相应元件的故障概率，采用基于 RMSprop 的 BP 方法对 DNN 进行训练；进而，针对 DNN 训练要求较大样本数而元件故障历史记录往往较少这一问题，本文给出了一种记录扩充提取方法，即同一变电站(或附近地区)同一类型的设备，因均经过严格的入网试验，其运行的微环境也是类似的，故可以把其历史故障记录看作相同或类似的，从而加大历史记录的数量；在此基础上采用基于概率统计和随机抽样的样本生成方法以产生足够的样本，从而实现模型的成功训练．在算例验证和分析部分，首先在模拟样本上验证了本文模型输出元件故障概率的正确性；进而，基于实际案例将本文方法与专家系统方法及浅层(单隐层)神经网络模型故障诊断方法进行了对比，结果表明本文方法相较于浅层神经网络方法，训练时间缩短了约 7%(4.05 s 缩短至 3.77 s)，测试误差减小了约 33%(由 0.005 1 减小至 0.003 4)；相较于专家系统，误辨率由 31.25%减小至 0．因此，本文的方法具有一定的潜力应用于实际电力系统，从而提升其诊断的正确率．
Abstract_English:As one of the standard modules of SCADA，the fault diagnosis software still has the problem of high misrecognition rates. To address this issue，this study proposes a new fault diagnosis model and algorithm based on the logistic regression deep neural network to improve accuracy. The building process of the new model and algorithm is as follows. First，a regression deep neural network (DNN)is established for each equipment，with fault feature vector whose elements have all been transformed into 1 or ?1 as the input，the fault probability of the corresponding equipment as the output，and the backpropagation method based on RMSprop as the training method. Then，to eliminate the difficulty caused by the need for a large sample size for training the DNN and the lack of actual fault history records，a new method is presented in this study to expand and achieve the required number of records for specific equipment. Given the fact that each type of equipment and its microenvironment have passed the rigorous tests before being operated，the record expanding method treats the historical fault records of the same type of equipment at the same substation or a nearby substation as the same or analogous，considerably increasing the number of historical records. On this basis，a sufficient sample size can be produced by the probability，statistical，and random sampling techniques and the model can be successfully trained. In the sample verification and analysis section，first，the correctness of the fault probability generated by the model is thoroughly verified using the simulation system. Then，a comprehensive comparison of the proposed method，the method based on the expert system，and the method based on the shallow neural network (one hidden layer)using the cases from practical power systems is performed. Results show that the proposed method reduces the training time by approximately 7% (i.e.，from 4.05s to 3.77s)，the test error by approximately 33% (i.e.，from 0.0051 to 0.0034)compared with the method based on the shallow neural network，and the misrecognition rate from 31.25% to approximately 0 compared with the method based on the expert system. Because of its easy implementation and high accuracy，the proposed method has considerable potential for application in practical power systems and can increase the accuracy of fault diagnosis.
Keyword_Chinese:故障诊断；深度学习；神经网络；故障概率
Keywords_English:fault diagnosis；deep learning；neural network；fault probability

PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6590