High Impedance Fault Detection Based on Attention Mechanism and Object Identification

Detection of high impedance faults (HIFs) has been traditionally a main challenge in the protection of distribution systems, since they do not generate enough current to be reliably detected by conventional over-current relays. Data-based methods are alternative HIF detection methods which avoid threshold settings by training a classification or regression model. However, most of them lack interpretability and are not compatible with various distribution networks. This paper proposes an object detection-based HIF detection method, which has higher visualization and can be easily applied to different scenarios. First, based on the analysis of HIFs, a Butterworth band-pass filter is designed for HIF harmonic feature extraction. Subsequently, based on the synchronized data provided by distribution-level phasor measurement units, global HIF feature gray-scale images are formed through combining the topology information of the distribution network. To further enhance the feature information, a locally excitatory globally inhibitory oscillator region attention mechanism (LEGIO-RAM) is proposed to highlight the critical feature regions and inhibit useless and fake information. Finally, an object detection network based You Only Look Once (YOLO) v2 is established to achieve fast HIF detection and section location. The obtained results from the simulation of the proposed approach on three different distribution networks and one realistic distribution network verify that the proposed method is highly effective in terms of reliability and generalization.