Real-time detection of short-circuit faults in power systems using LSTM autoencoder

This paper presents a real-time detection method for short-circuit faults in meshed power networks, based on a Long Short-Term Memory (LSTM) autoencoder. The objective is to quickly and accurately identify both simple faults (single-phase, double-phase, three-phase), hybrid faults, and triple faults characterized by the simultaneous occurrence of several types of faults on different network lines. Simulations were carried out on an IEEE 9-bus system modeled in MATLAB/Simulink, covering various scenarios: faults on a single line, on two lines, and on three lines simultaneously. The results obtained show an immediate reaction of the model to disturbances, reflected by a marked increase in the loss function at the time of the incident. The system also demonstrated its ability to detect the indirect impact of faults on neighboring lines, reflecting the propagation of electrical imbalances throughout the network. These performances confirm the effectiveness and robustness of the LSTM autoencoder for intelligent, fine, and adaptive monitoring of modern power networks, fully meeting the requirements of real-time protection systems.