Real-Time Fault Detection in Power Systems Using Adaptive Anomaly Detection

Abstract

Anomaly detection in current and voltage data is critical to ensuring the reliability and safety of electrical systems. This research presents a comprehensive framework that combines feature extraction and machine learning to identify irregularities in time series measurements. The study evaluates five anomaly detection algorithms, namely Z Score, Spectrum Detector, IsolationForest Detector, DeepSVDD, and AutoEncoder, using precision, recall, and F1 score metrics. A systematic feature extraction pipeline, including time frequency analysis, entropy measurement, and signal coherence, is employed to improve detection accuracy. The results show that DeepSVDD achieves the highest precision (0.85), while AutoEncoder excels in recall (0.90), highlighting their respective strengths in minimizing false positives and maximizing anomaly coverage. The F1 score analysis reveals that DeepSVDD and AutoEncoder are the most balanced performers, offering robust solutions for diverse operational requirements. This work provides empirical insights into algorithm selection for industrial applications, bridging the gap between theoretical methods and practical implementation in power systems.