Arc fault detection using LSTM autoencoders and isolation forest in electrical distribution networks

The presence of arc faults in medium voltage distribution system presents substantial hazard to the reliability and security of the system, owing to the possibility of electrical fires and equipment damage. The random and unpredictable nature of arc faults sometimes poses challenges for traditional arc fault detection systems, thereby requiring the development of more advanced approaches. This work introduces an approach for detecting arc faults by integrating a Long Short-Term Memory (LSTM) autoencoder model with Isolation Forest (IF). The proposed approach aims to learn and classify normal operational patterns while efficiently identifying abnormalities resulting from arc faults. The LSTM autoencoder model accurately reconstructs normal time-series data, detecting substantial variations in reconstruction that are possible faults. The reconstruction errors are further analysed using IF to effectively differentiate abnormal behaviours from normal operations. This study assesses the proposed method's efficacy in detecting arc faults by analysing current time series data from a medium voltage distribution system. The results demonstrate Mean Squared Error (MSE) of 0.0068, Mean Absolute Error (MAE) of 0.0619, and R-Squared (R²) of 0.9369 in arc fault identification, providing a viable solution for improving system reliability and reducing arc fault related occurrences in medium voltage networks.