Performance Comparison Between Deep Learning Models for Fault Classification in Transmission Lines Using Time Series Data

Abstract

Deep learning has become a vital tool for addressing complex challenges in power systems, particularly fault detection and classification in transmission lines. This study presents a comparative analysis of three advanced time-series models like temporal convolutional networks (TCN), bidirectional long short-term memory (BiLSTM), and gated recurrent units (GRU) for fault classification. Leveraging a comprehensive data set encompassing diverse fault scenarios like single-phase to ground fault (AG), double line to ground fault (ABG), three-phase fault (ABC) from both simulated and real transmission line data, the study provides a rigorous evaluation of these models’ performance under realistic conditions. The results demonstrate that TCN achieves a fault classification accuracy of 99.9%, significantly outperforming BiLSTM (92.31%) and GRU (95.27%), while also excelling in precision, recall, F1 score, and training efficiency. Additionally, this study incorporates feature extraction techniques like discrete wavelet transform (CWT) to establish new benchmarks for fault classification. The findings underscore TCN's robustness in handling the dynamic nature of transmission line signals and its practical potential for real-time applications, contributing to the development of more reliable and efficient power system fault classification systems.