A deep learning based intelligent approach in detection and classification of transmission line faults

Abstract

Detection and classification of transmission line (TL) faults are key factors for the fault root cause analysis and rapid restoration of the power network. Deep learning can extract representative features automatically from big data, thus becoming an important breakthrough in fault analysis. Existing deep learning-based fault diagnosis models rely on a large number of data obtained from a variety of fault conditions to be generalized. Indeed, in the TL domain, it is difficult to define and, therefore to collect the fault incepted features of all possible fault scenarios. This paper proposes an unsupervised framework for fault detection and classification (FDC) of TL based on a capsule network (CN). Instead of using the baseline CN, an extension to this with a sparse filtering technique is adopted in this work. The capsule network with sparse filtering (CNSF) voluntarily learns the expensive fault features and significantly improves the model performance without involving a large number of data. The proposed scheme receives 1/2 cycle post-fault three-phase signals and encodes them into a single image that is defined as the input for the proposed CNSF model. The effectiveness of the proposed CNSF model is corroborated by four different TL topology confirming the model’s adaptability to a topology change in response to intended control action or the switching actions due to cascading faults. Further assessment of the model’s performance against noise, high impedance faults (HIF), and line parameter variations are also carried out in order to confirm the high reliability of the proposed model. In addition, a rigorous comparative study is conducted to guarantee the state-of-the-art performance of the proposed model.