Transactions of China Electrotechnical Society  2023, Vol. 38 Issue (8): 2178-2190    DOI: 10.19595/j.cnki.1000-6753.tces.211990
Current Issue| Next Issue| Archive| Adv Search |
Fault Line Diagnosis Scheme of Active Distribution System Based on Time-Sequence Convolution Residual Network
Chu Xu1, Bao Zehong1, Xu Liqiang2, Yan Yabing2
1. College of Electrical and Information Engineering Hunan University Changsha 410082 China;
2. Electric Power Research Institute of State Grid Hunan Electric Power Co. Ltd Changsha 410007 China

Download: PDF (2133 KB)   HTML (1 KB) 
Export: BibTeX | EndNote (RIS)      
Abstract  The fault characteristics of active distribution system (ADS) are weak due to its complex structure and flexible control strategy, so the diagnosis method with strong adaptability and high precision is urgently needed. This paper examines a new fault diagnosis scheme for lines of active power distribution system founded on a time-sequence convolution residual network, which is used to detect, select and locate short circuit fault. According to the strong correlation between each feature channel signals measured at corresponding time of pre-fault and post-fault period, a residual network model based on time-sequence convolution (T-Conv) is built. For the proposed T-Conv residual network model, each time-sequence convolution kernel transverse telescopic width based on the number of adaptive features matching the input channel, ensure the receptive field lateral coverage to all the characteristics of the channel, to fully mix the channel between the features of information, enhance the receptive field, reducing the loss of information, and save the calculation process of convolution kernels lateral translation, save calculation time and improve calculation efficiency. The kernel size and step size are set according to the temporal morphological characteristics of fault transient and sampling frequency, and convolved downward along the time axis. The fault local features on the time axis are successively mined, mutation details are extracted from multiple directions, and complex, abstract and weak features of ADS fault signals are extracted efficiently. Furthermore, the short nature of residual network is used to generate a backoff mechanism to solve the problem of gradient disappearance in deep neural networks. Thus, each network layer can obtain enough gradients to update its parameters, which reduces the difficulty of model training. Fault selection and fault location belong to both classification tasks, and the front part of the network belongs to feature extraction. After mining corresponding features through training, only the corresponding number of neurons are needed to output the discriminant probability of each category according to the different task objectives. Therefore, this paper realizes fault selection and fault location functions by modifying the structure of the output layer of the model. In addition, the fault detection function can be realized and the reliability of fault diagnosis can be further improved by considering the normal category and combining with Boolean operator in the selection and location discrimination results. The fault diagnosis scheme results guide the staff to find abnormalities and repair faults in time, accelerate system recovery, and form an active fault diagnosis scheme for the distribution system that integrates fault detection, selection, and location functions. Finally, an electromagnetic transient model of 10 kV active distribution system is built using EMTDC. Experimental results show that for different transition resistance, fault location and fault initial Angle, this article proposed fault diagnosis scheme based on time-sequence convolution were diagnosed with high precision, fast computing speed, the advantages of without expert experience, avoid working condition of fault signal feature selection, time and frequency domain transformation, such as complicated process, direct "end-to-end" build a diagnosis model, implement exemption threshold diagnosis. Compared with other similar diagnosis methods based on recurrent neural network (RNN), artiflcal neutral networks (ANN) and convolutional neutral networks (CNN), which have single function and do not consider multi-function combination, this method effectively integrates and realizes the functions of fault detection, fault selection and fault location, and has advantages in diagnosis accuracy and calculation efficiency. In addition, as the proportion of training set samples decreases, the model still has high diagnostic accuracy, showing adaptability to different sample sizes and good generalization ability.
Key wordsActive distribution system      fault diagnosis      time-sequence convolution      residual network     
Received: 10 December 2021     
PACS: TM75  
Service
E-mail this article
Add to my bookshelf
Add to citation manager
E-mail Alert
RSS
Articles by authors
Chu Xu
Bao Zehong
Xu Liqiang
Yan Yabing
Cite this article:   
Chu Xu,Bao Zehong,Xu Liqiang等. Fault Line Diagnosis Scheme of Active Distribution System Based on Time-Sequence Convolution Residual Network[J]. Transactions of China Electrotechnical Society, 2023, 38(8): 2178-2190.
URL:  
https://dgjsxb.ces-transaction.com/EN/10.19595/j.cnki.1000-6753.tces.211990     OR     https://dgjsxb.ces-transaction.com/EN/Y2023/V38/I8/2178
Copyright © Transactions of China Electrotechnical Society
Supported by: Beijing Magtech