Deep Learning Based High Impedance Fault Detection in Electrical Distribution System Using CNN–BiLSTM with Noise Robustness Analysis

Abstract

High impedance faults (HIFs) in electrical distribution systems are difficult to detect due to their low fault current magnitude and irregular behaviour. The growing integration of distributed energy resources further increases system complexity, requiring more reliable and adaptive fault detection methods. Traditional methods of protection typically do not find HIFs because the currents that flow through the fault are similar to the normal load current and do not activate the conventional relays (that use fault current levels). In addition, the existing approaches for signal processing and machine learning techniques, which are commonly used to detect HIFs, are very sensitive to noise and require threshold-value settings to be successful, leaving them with limited capabilities to capture non-linear characteristics of the fault, resulting in decreased effectiveness as system operating conditions change. This research presents a hybrid deep-learning architecture that combines the strengths of both Convolutional Neural Networks (CNNs) to extract spatial features and Bidirectional Long Short-Term Memory (BiLSTM) networks to learn the temporal relationships in fault behaviours. The CNN is designed to extract spatial features from voltage and current signals. The BiLSTM is designed to learn temporal dependencies 
between faults for detecting faults in real use cases. A comprehensive preprocessing pipeline is developed that includes feature construction, normalisation, class balancing and noise augmentation to improve the generalisation of the model. With an overall accuracy of 94.92 %, the model is reliable in accurately detecting faults as well as robust against noisy data, making it ideal for use in real-time smart grid protection applications.