An extensive analysis of machine learning algorithms for the identification and differentiation of high impedance faults in various power network topologies

This is the authors' abstract. We don't add key points for this paper.

The traditional protection systems, which are usually deployed in the electrical network, are unable to detect High Impedance Faults (HIFs) because of their low amplitude and non-linear characteristics. In this paper, innovative HIF analysis techniques using machine learning (ML) and deep learning (DL) methods for various power system types, i.e., radial systems, IEEE test systems, and microgrid networks, are discussed. Through a data synthesis of 101 algorithms and 50 studies, it established that transformer-based DL solutions have shown superior results with an average of 99.3%, followed by 98.2% in hybrid solutions, 98.0% in Convolutional neural network (CNN)-based DL solutions, Recurrent neural network (RNN)/Long short-term Memory (LSTM)/Gated recurrent unit (GRU) solutions with 97.8%, and 97.2% in tree-based solutions. Shallow ML algorithms have a minimum accuracy of 96.5%. The most widely used algorithms are CNNs (35%) and transformers (19%). The study identified critical gaps: 78% simulation, 22% real-world tests; 38% of the studies ignore data imbalance; and interoperability among microgrid clusters is not addressed. This research gives utilities and protection engineers analytical validation to choose relay firmware for transformer hybrid systems, achieving over 99% accuracy in new deployments, and supports cost-efficient shallow ML solutions for existing infrastructure, achieving over 96.5% accuracy in modernisation.