Adaptive fault diagnosis in power transmission lines using deep learning and LSTM autoencoders for enhancing grid reliability

Abstract

The increasing complexity and dynamic nature of modern electrical grids necessitate advanced, adaptive fault diagnosis systems to maintain high reliability and ensure minimal downtime. This study presents a novel, adaptive fault detection and localization method for three-phase transmission lines utilizing a Long Short-Term Memory (LSTM) Autoencoder. The model operates in an unsupervised manner, learning the standard operational patterns from three-phase voltage and current signals and identifying faults as anomalies through high reconstruction errors. Trained and tested on a comprehensive dataset of over 50,000 simulated fault events generated in MATLAB/Simulink and rigorously validated on 1,000 real-world fault instances from an open-source repository, the proposed method demonstrates exceptional performance and robustness. It achieves a 98 % accuracy and a 2 % false positive rate, outperforming traditional methods (DFT, Wavelet) and other deep learning benchmarks (standard LSTM, 1D-CNN) by over 15 % in F1-score. The model exhibits strong resilience to noise, maintaining an F1-score above 92 % at a 20 dB SNR, and demonstrates computational efficiency suitable for real-time deployment. These results validate the LSTM Autoencoder as a potent and practically significant tool for enhancing adaptive fault management and real-time monitoring in modern power systems, directly contributing to improved grid reliability and operational efficiency.