A protection scheme based on conditional generative adversarial network and convolutional classifier for high impedance fault detection in distribution networks

• Published in: Electric power systems research 2022-11, Vol.212, p.108633, Article 108633
• Main Authors: Mohammadi, Amin, Jannati, Mohsen, Shams, Mohammadreza
• Format: Article
• Language: English
• Subjects: Conditional generative adversarial network and convolutional classifier; Distribution network; Fault detection; High impedance fault
• ISSN: 0378-7796; 1873-2046
• DOI: 10.1016/j.epsr.2022.108633

•Combination of Conditional GAN with Convolutional classifier is proposed for fault detection.•Third Harmonic Angel (THA) of fault current is used as an effective feature.•Little amount of training data is used to train CGAN.•High performance of the CGAN in generation of pseudo-real data in large scale. Low level current and similarity of High Impedance Faults (HIF) in respect of characteristics to other transient events have posed a critical challenge to the protection of distribution systems. In addition, the dependency of previous methods on large amounts of training data increases the simulation error rate, and preparing this amount of data is time-consuming. In this paper, a novel scheme based on conditional generative adversarial network (CGAN) and convolutional neural network (CNN) classifier techniques is proposed, that reduces this dependency and leads to acceptable classification accuracy. In the proposed method, a small amount of data is extracted from the under-study network as the real data. Then, the third harmonic angle of the current is extracted from the real data by an adaptive linear neuron (ADALINE) as an effective feature. The CGAN is performed to produce a large amount of pseudo data. At last, the fault data is separated from other transient network events via the CNN classifier. Five different scenarios are used to evaluate the proposed method on a 13-bus IEEE network. The simulation results show that the Precision and Recall of distinguishing HIFs from other transient events is greater than 98% in all the scenarios. These results verify that the proposed scheme is very accurate despite the low dependency on input training data. [Display omitted]