Umeå University's logo

Ground fault relays (GFRs) in resonant-grounded medium voltage distribution networks shall not operate during phase-to-ground (Ph-G) fault inception, allowing the Petersen coil to suppress self-extinguishing faults, but the designated GFR must operate during permanent faults. In order to enhance the performance of GFRs, particularly during high-impedance faults, the scope of this paper is to propose a straightforward, machine-learning-based protection framework. The enhanced GFR is modeled as a classification task. Depending on the GFR's position and the Ph-G fault location in the network, fault samples are labeled as “no operation,” “primary,” “backup,” or “backup of backup,” forming two-class, three-class, and four-class GFR setups, respectively. This assures selective operation across three protection zones and improves the reliability of all GFRs. The proposed protection scheme employs backward optimal feature selection to identify the most relevant discrete features obtained from measured zero-sequence current and voltage waveforms. An ensemble of k-nearest neighbor classifiers is utilized for accurate classification, simulating the GFR operating conditions, with measurement errors and sensitivity incorporated in the preprocessing. A 20 kV case study network validates the proposed framework, achieving F1-scores exceeding 96 %. The maximum operation delay of the protection scheme for an enhanced GFR is 225 ms, accommodating the required time window (200 ms), prediction time (5 ms), and change detection time (20 ms), thus assuring safe operation. Compared to other machine-learning-based methods used for Ph-G fault protection in resonant-grounded radial networks, this framework is high-performing, fast, and easy to implement, utilizing a simpler structure than neural networks.