Data-driven Earth-Fault Localization via Neural Network Models for Classification and Regression

Rapid detection and precise localization of earth
faults are crucial for ensuring the safe operation of energy
distribution grids. Although detection is generally accom
plished using integrated power quality hardware, pinpointing
the fault location frequently requires labor-intensive manual
switching operations, which can lead to supply disruptions.
This work explores two methods of automated, robust, and
accurate earth fault localization in medium voltage power grids
using machine learning techniques. A digital twin is utilized to
simulate fault events thus enabling the generation of extensive
data to facilitate the training of data-driven models for fault
localization. We explore two shallow neural network models
to tackle the fault localization problem. On the one hand, we
predict potential fault locations via node classification. On the
other hand, a regression model is realized, that maps input
measurements to associated impedances.
This research work considers real-world medium voltage
grids and focuses on investigating the influence on accuracy,
robustness and flexibility of the implemented fault localization
methods. The aim is to characterize the application areas of
each approach for distribution grid operators. Finally, we open
a discussion on requirements and desirable improvements for
future developments.