Explainable Machine Learning for Resource-Efficient Fault Classification in Power Grid Oscillograms: Impact of Measurement Channel Selection
Year: 2026
Type: conference-paper
Abstract: Reliable fault classification in power distribution networks is essential for relay protection, grid stability, and automated fault management, yet the operational impact of measurement channel selection on classification performance remains poorly understood. This paper presents a systematic featuregroup ablation study using the OscGrid dataset, a publicly available collection of$\mathbf{5 0, 7 6 5}$real-world oscillograms from$\mathbf{0. 4}$−35 kV substations, to quantify how channel availability shapes both model performance and data volume. Three channel configurations are evaluated: phase currents only (Set A), currents with busbar voltages (Set B), and the full set including cable line voltages (Set C). Random Forest (RF), Support Vector Machine (SVM), and Gradient Boosted Trees (BT) are compared under stratified 5-fold cross-validation with balanced class weights on a binary fault classification task derived from the$\mathbf{M L}_{3}$taxonomy. Busbar voltages emerge as the dominant discriminative group, with RF F1 increasing from 0.534 (Set A) to 0.858 (Set B), corresponding to a gain of +0.324. Adding cable line voltages yields the best performance$(\mathbf{R F} \mathbf{F} \mathbf{1} \boldsymbol{=} \mathbf{0. 9 3 1}, \mathbf{A U C} \boldsymbol{=} \mathbf{0. 9 9 5})$but reduces usable samples by 49.7 percent (from 77,756 to 39,094), revealing a critical trade-off between feature richness and data availability that is not captured by accuracy metrics alone. Mean Decrease Impurity analysis consistently ranks busbar voltages (UABB, UBBB, UCBB) as the most influential channels, consistent with the voltage asymmetry signatures of ground fault events, while cable line voltages contribute less than half the importance of busbar channels. These findings provide engineeringlevel, measurement-aware guidance for deploying automated fault detection in real substation environments. These findings also highlight the importance of resource-efficient measurement strategies in partially instrumented power systems, supporting more sustainable and cost-effective monitoring infrastructures.
Author: Ezgi Guney
Institution: Sinop University
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203596899