Search

Although the use of a series capacitor increases the transmission line (TL) capacity, it inverts the voltage/current and creates a sub-harmonic frequency that results in unintended operation of the protection devices, which further decreases the transient stability under fault conditions. Restoring the transient stability requires an accurate diagnosis of faults in TLs, which further requires exploration of in-depth features. This imposes challenges for the existing approaches as they demand a large amount of data. In this paper, a self-attentive weight-sharing capsule network (WSCN) is proposed to achieve robust and high classification performance while utilizing a limited amount of data. The capsule network, with the weight-sharing mechanism delivers robust performance in detecting and classifying the faults in the TL domain. Simultaneously, the self-attention layer highlights the more dominant features that make the network work upon the limited data effectively. A Western-System-Coordinating-Council WSCC 9-bus and 3-machine test model, which was modified with the series capacitor was studied to quantify the robustness of the self-attention WSCN. The results were compared with those of others networks to confirm the explicit classification ability of the proposed network. A study on real-world data obtained from power network was also conducted to validate the proposed self-attention WSCN. [ABSTRACT FROM AUTHOR]

• Proposed an intelligent approach for transmission line fault analysis based on capsule network. • An unsupervised model is developed which requires no manual labeling of the data. • The performance against line parameter variation and high impedance faults is assessed. • Noises are considered to justify the robustness of the proposed approach. • The performance is compared with the baseline capsule network and existing approaches. 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. [ABSTRACT FROM AUTHOR]