Fault detection and classification of bipolar DC system with dedicated metallic return based on TF-ENSR

Fast and accurate fault diagnosis is necessary for continuous power delivery. In bipolar DC systems with dedicated metallic return (DMR), accurately distinguishing between pole-to-ground and pole-to-DMR faults is challenging. To address this, this paper proposes a data-driven protection method based on feature learning. Specifically, time-series feature extraction based on scalable hypothesis tests (Tsfresh) is utilized to automatically extract the feature set with clear physical interpretations from fault voltage signals. To ensure that the feature set is sufficiently concise while fully describing the fault state, the Feature-selector is integrated to efficiently eliminate redundant features. Additionally, an improved elastic network softmax regression (ENSR) model is employed to establish an accurate mapping between the selected features and fault types. Offline simulations in PSCAD/EMTDC and hardware-in-the-loop testing with real-time digital simulators (RTDS) demonstrate that the proposed method effectively detects and distinguishes between fault types. Comparative studies with other protection methods highlight the advantages of the proposed approach, including its robustness to high transition resistances and strong noise interference, adaptability to changes in sampling environments and current-limiting reactors, and low online computational burden.