Time-Series vs. Statistical Methods for Small Current Fault Detection: A Comparative Review of Accuracy and Efficiency

Abstract

Small current grounding faults pose persistent challenges in power systems due to their low-magnitude signatures and the complexity of timely, accurate detection. While deep learning methods have recently gained attention, traditional time-series and statistical techniques remain foundational in industrial applications for their interpretability, computational simplicity, and compatibility with legacy infrastructure. This paper presents a structured review of non-deep learning approaches for detecting small current faults using two-channel current-voltage measurements, with a focus on balancing detection accuracy and computational efficiency. We analyze 12 influential studies from 2019–2024, comparing methods such as wavelet transforms, Hilbert-Huang analysis, RMS-based detection, and hybrid approaches. A key contribution is the introduction of the "Detection Intelligence vs. Implementation Agility" framework, which refines the selection paradigm beyond basic performance trade-offs. Results from our analysis and a supporting comparative simulation show that advanced time-series techniques achieve accuracies exceeding 95% but demand considerable processing resources, while statistical methods offer sub-20-ms response times with reduced sensitivity in noisy conditions. Hybrid methods bridge this gap but face complexity barriers. This review contributes a decision framework for method selection based on system requirements, identifies gaps in benchmarking, and outlines future directions, including the need for explainable and lightweight hybrid algorithms.