Understanding the limits of snapshot-based fault diagnosis in distribution networks

Ba Tu Phung, Huy Hoang Le, Kim-Anh NguyenarXiv 2026 · arXiv

TL;DR

Snapshot-based fault diagnosis—classifying grid faults using a single "instant" of voltage data—is a highly effective, low-bandwidth screening tool for modern distribution networks. This paper reveals that the performance bottleneck isn't the machine learning model, but rather measurement noise on high-impedance faults; however, using just three well-placed sensors can achieve up to 99% accuracy when paired with high-precision micro-PMUs.

What Problem Does This Paper Address?

As distribution networks integrate more inverter-based resources (IBRs) like solar PV and battery storage, traditional fault protection becomes significantly more difficult. Unlike synchronous generators, IBRs use controllers that limit fault currents, making faults look "weaker" and harder to distinguish from normal load variations.

"Snapshot-based" diagnosis offers a computationally cheap solution by analyzing voltage phasors at a few key locations at a single moment in time. While many studies have reported high accuracies for this approach, they often use inconsistent noise assumptions and fail to explain the root cause of errors. Are we limited by the classifier's architecture, the features we extract, the sensor noise, or the physical resistance of the fault itself? This paper systematically decomposes these error sources to find the true "operating envelope" of snapshot diagnosis.

Methods

The authors developed a rigorous evaluation framework (shown in Figure 2) that moves beyond simple accuracy scores:

  • Error Decomposition: They tested 13 different classifiers (including LightGBM, Random Forests, and CNNs) across four feature sets of increasing complexity—from simple aggregate statistics (AGG) to the full phasor snapshot (FPS).
  • Standardized Noise Modeling: Instead of arbitrary signal-to-noise ratios, they used the Total Vector Error (TVE) classes defined by IEEE standards. This allows researchers to relate model performance directly to real-world hardware, such as standard 1% TVE sensors versus high-precision micro-PMUs (0.02% TVE).
  • Discriminability-Margin Placement (DMP): A novel, fast sensor placement method. It uses Integer Linear Programming (ILP) to select sensor locations based on simulations alone, maximizing the "margin" between different fault types in the presence of noise.
  • The "Reject Option": An abstention mechanism that allows the model to flag uncertain snapshots (typically high-impedance faults) for more detailed, slower secondary analysis rather than providing a potentially wrong guess.
Figure 2: Workflow of data generation, sensor placement, measurement and evaluation.

Datasets

The study utilized a mix of standard benchmarks and large-scale synthetic data:

  • IEEE Test Feeders: The 13-node, 34-node, and 123-node radial feeders, simulated using OpenDSS to capture unbalanced phases and network asymmetries.
  • DER-Fault Dataset: A public dataset featuring 12,600 fault events on an Iowa 240-bus system with varying levels of distributed energy resource (DER) penetration.

Key Insights and Ideas

  • The Classifier is Not the Bottleneck: One of the most striking findings is that with "perfect" (noise-free) data, the full phasor snapshot classifies almost every fault correctly. This means researchers should spend less time tweaking neural network architectures and more time improving data quality and sensor placement.
  • The High-Impedance Wall: Accuracy drops sharply as fault resistance ($R_f$) increases. At the 1% TVE noise level, faults above 10–32 $\Omega$ produce voltage changes comparable to the sensor's own noise, creating a physical limit on detectability.
  • Hardware Over Algorithms: Upgrading from a standard 1% TVE sensor to a micro-PMU provides a much larger accuracy boost (up to 9 points) than switching between different machine learning models.
  • Placement Matters (But Only a Little): While the DMP method is incredibly fast (selecting sensors in under a second), the paper finds that snapshot measurements are highly redundant. On smaller feeders, even "hand-picked" or random sensor sets perform reasonably well, though DMP excels at catching difficult high-impedance faults.

Key Results

  • Sparsity: Just three voltage sensors are sufficient to classify 90–96% of faults at the 1% accuracy class. With micro-PMUs, this jumps to 96–99%.
  • Speed: All tested classifiers completed inference in under 25 ms, making them ideal for real-time "first-stage" screening.
  • Reliability: By using the "reject option" to defer the most uncertain 4–16% of cases (mostly high-impedance faults), the system's accuracy on the remaining snapshots rises to 99%.
  • Generalization: The sequence-component features (SEQ) proved the most robust when transferring a model trained on a grid without DERs to one with high solar penetration.
Table 1: Comparison with recent data-driven fault diagnosis studies in distribution networks.

Limitations and Future Work

The authors acknowledge several limitations that define the next frontier for this research:

  • Steady-State Assumption: The models use steady-state phasor data. In reality, the timing of the snapshot relative to fault transients could introduce further errors.
  • Voltage Only: The study focused exclusively on voltage phasors because they are cheaper to measure. Adding current measurements at the feeder head could likely bypass the "attenuation" issues seen at the substation.
  • Field Validation: While the study used high-fidelity simulations and public datasets, validation on real-world field recordings remains a necessary next step to account for unmodeled grid dynamics.

Summary updated

Abstract

Inverter-based resources weaken fault currents in distribution networks, and snapshot-based diagnosis, which classifies a fault from voltage phasors at a few buses at one instant, offers a cheap first-stage screen. Its published accuracies, however, rest on different noise assumptions and do not show whether the error comes from the classifier, the features, the measurement error, or the fault resistance. This paper separates these sources on three IEEE test feeders and a public data set with distributed generation, using four feature sets, thirteen classifiers, and measurement error defined by total vector error classes. The classifier is not the limiting factor, because the full phasor snapshot classifies every fault correctly without noise, and the remaining error comes from measurement error on high-impedance faults. Three voltage sensors classify 90-96% of faults at the 1% accuracy class and 96-99% with distribution micro phasor measurement units (micro-PMUs). A discriminability-margin placement selects them from simulations alone in under one second and matches the best of twelve placement methods. The results show when snapshot-based diagnosis can be trusted.
BibTeX
@misc{phung2026understanding,
  author = {Ba Tu Phung and Huy Hoang Le and Kim-Anh Nguyen},
  title = {{Understanding the limits of snapshot-based fault diagnosis in distribution networks}},
  year = {2026},
  eprint = {2611.17998},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.17998},
}
Figure