Hybrid model of convolutional auto-encoder and ellipse characteristic for unsupervised high impedance fault detection

High impedance faults (HIFs) present significant safety hazards such as fires and electric shock accidents, urging the development of reliable HIF detection approaches. Unfortunately, conventional protection relays often fail to detect HIFs due to their low fault currents. The growing utilization of sensors in distribution grids enables the monitoring of node signals, offering a promising way for accurate HIF detection. Nevertheless, challenges persist due to the scarcity of faulty data and the high costs associated with data collection. To tackle this challenge, we propose a hybrid artificial intelligence (AI) approach that combines the unsupervised learning capabilities of auto-encoders with the reliable physical knowledge of voltage waveform. This approach involves the transformation of node voltage signals into phase diagram curves, leveraging the ellipse characteristics of the curve to normalize the model’s learning process. Performance evaluation conducted using the IEEE-34 and IEEE 13 nodes test feeders demonstrates its effectiveness in achieving highly accurate HIF detection across diverse fault conditions and locations. Comparative analysis against other unsupervised HIF detection methods reveals its superiority in terms of high detection accuracy and strong robustness to switching disturbances. The simulation data and codes relevant to this study are openly accessible, serving as a benchmark for research in HIF detection. • Developing phase diagrams curve and ellipse loss of voltage signal. • Combining a convolutional auto-encoder with ellipse constraints for feature extraction. • Integrating the ellipse loss with the reconstruction loss for HIF detection. • Validating the superiority of the proposed model via two simulation models. • Making simulation data and related code publicly available.