Harmonic Signature-Based One-Class Classifier for Islanding Detection in Microgrids

This article presents a new passive islanding detection technique in MGs that uses locally measured voltage signals at the PoC of DERs. The proposed method distinguishes islanding events from normal/non-islanding conditions by utilizing superimposed harmonic spectra extracted through a full-cycle discrete Fourier transform. Our solution utilizes a machine-learning-based one-class classifier to define and adjust thresholds for full harmonic spectra. Unlike other methods, our approach does not require data synchronization or communication infrastructure, nor does it suffer from common errors that often arise in current transformers. Moreover, our design is compatible with distributed and decentralized control strategies, as it relies solely on local voltage measurements at the PoC. Another advantage of this method is its low sampling frequency requirement, in the range of 1 kHz, making it cost-effective and implementable in most existing systems. In a comprehensive evaluation of a typical MG test system that included synchronous and inverter-based DERs, the proposed scheme demonstrated exceptional performance. Specifically, the scheme was able to detect 99.06% of different islanding events within the training range, with a detection time of just 10 to 21 ms. Additionally, the scheme remained 100% stable during various normal conditions, short-circuit faults, load changes, voltage changes, capacitor switching, and frequency changes.