Nicolay V. Khripunov,¹ Olga A. Filippova,² Vladimir A. Berdnikov,³ Nataliya А. Sosina,¹ Oksana V. Oskina<sup>1

1</sup>Togliatti State University (Russian Federation)
²Volga Region State University of Service (Russian Federation)
³Togliatti Academy of Management (Russian Federation)

Abstract

This paper analyzes the application of machine learning methods for fault localization and sectionalization recommendations in a distribution network with a high penetration of inverter-based sources. The study investigated the use of a lightweight graph neural network for simultaneous classification of fault location and prediction of switching control actions. A procedure for constructing a synthetic radial feeder and generating scenarios with noise, telemetry dropouts, and a variable proportion of distributed generators is described. The application of a linear voltage distribution model for rapid calculation of node features during synthetic simulation was studied. A set of input features for the graph neural network was formed for each node, including voltages, currents, active and reactive powers, and inverter-based generator status flags. A graph neural network architecture was developed, comprising two message-passing layers, a global readout, and two output layers for fault localization and action recommendation tasks. A training dataset of scenarios was formed in Google Colab, divided into train/val/test sets, with variations in noise levels and measurement incompleteness. Experiment support was developed, including data generation scripts, model training procedures, a set of metrics for localization and recommendation accuracy, and inference time measurements. A comparison with basic threshold detection and an analysis of performance degradation with increasing inverter-based source penetration and data dropouts were justified. Recommendations for future model expansion directions were developed.