A Robust Self-Attentive Capsule Network for Fault Diagnosis of Series-Compensated Transmission Line

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.

Automated summary

The proposed self-attentive weight-sharing capsule network (WSCN) achieves robust and high classification performance in detecting and classifying faults in the TL domain, utilizing a limited amount of data. By utilizing a weight-sharing mechanism and self-attention layer, the network effectively highlights dominant features and demonstrates explicit classification ability, as confirmed through experiments on a modified test model and real-world data from power networks.