Data-Driven Linearization based Arc Fault Prediction in Medium Voltage Electrical Distribution System

TL;DR

This paper introduces a data-driven linearization framework for early detection of high-impedance arc faults in medium-voltage electrical systems, enabling predictions approximately 11 milliseconds before fault occurrence with high accuracy.

Contribution

The study presents a novel data-driven linearization method that transforms nonlinear waveforms into a linear space for early fault prediction, improving interpretability and prediction timeliness.

Findings

Successfully predicted arc faults 11 milliseconds before actual occurrence.

Achieved accurate fault precursor modeling using only pre-fault healthy data.

Demonstrated robustness through eigenvalue and waveform fidelity analyses.

Abstract

High-impedance arc faults (HIAFs) in medium-voltage electrical distribution systems are difficult to detect due to their low fault current levels and nonlinear transient behavior. Traditional detection algorithms generally struggle with predictions under dynamic waveform scenarios. This research provides our approach of using a unique data-driven linearization (DDL) framework for early prediction of HIAFs, giving both interpretability and scalability. The proposed method translates nonlinear current waveforms into a linearized space using coordinate embeddings and polynomial transformation, enabling precise modelling of fault precursors.The total duration of the test waveform is 0.5 seconds, within which the arc fault occurs between 0.2 seconds to 0.3 seconds. Our proposed approach using DDL, trained solely on the pre-fault healthy region (0.10 seconds to 0.18 seconds) effectively captures certain invisible fault precursors, to accurately predict the onset of fault at 0.189 seconds, which is approximately 0.011 seconds (i.e., 11 milliseconds) earlier than the actual fault occurrence. In particular, the framework predicts the start of arc faults at 0.189 seconds, significantly earlier of the actual fault incidence at 0.200 seconds, demonstrating substantial early warning capability. Performance evaluation comprises eigenvalue analysis, prediction error measures, error growth rate and waveform regeneration fidelity. Such early prediction proves that the model is capable of correctly foreseeing faults which is especially helpful in preventing real-world faults and accidents. It confirms that our proposed approach reliably predicts arc faults in medium-voltage power distribution systems