PaperlayerView exact revision on arXiv

Source-linked AI summary

A Topology-Aware Neural Network for Fault Location With Missing Measurements

Fan Jiang, Linhan Fang, Xingpeng Li, Kaushik Rajashekara

arXiv:2610.36474v1eess.SYcs.SY

TL;DR

Distribution-feeder fault location must distinguish among electrically similar locations, while measurement devices may not always be available. TANFL combines component identification, continuous line-position estimation, and missing-input handling; on the tested feeder, it outperforms the MLP and leads most averaged one-block-missing accuracy comparisons.

  • Problem

    Prior work had not combined continuous fault location, bus-versus-line identification, and robustness to missing inputs in a single model, despite feeder ambiguity and unavailable measurements.

    Source: §I · p. 1

  • Method

    TANFL uses post-transient voltage and current phasors, feeder connections, and line parameters to identify a faulty bus or line and estimate position on a selected line.

    Source: §I · p. 2

  • Results

    On the IEEE 15-bus test feeder, TANFL exceeded the MLP at 5% tolerance with complete and missing measurements, and led line accuracy at 3%, 5%, and 10% and bus accuracy at all tested tolerances across 11 one-block-missing scenarios.

    Source: §IV · p. 5

  • Takeaways & Limitations

    The results support TANFL’s effectiveness under the tested measurement conditions, including scenarios with one missing bus measurement block.

    Source: §IV · p. 5

    Evaluation is on the IEEE 15-bus feeder; extending the method to other fault types remains future work.

    Source: §IV · p. 5

Abstract

from arXiv · show

This paper proposes a topology-aware neural network-based fault locator (TANFL) for radial distribution feeders, which uses voltage and current phasors obtained after the decay of fault induced transients, as well as feeder connection relationships and line parameters, as inputs. The shared bus encoder, bus readout module, and gated line readout module work together to identify the faulty bus or line and estimate the continuous fault location on the selected line. During training, explicit missing data indicators and random bus data block masking are introduced, enabling a single trained model to handle both complete measurement sets and scenarios where any permitted bus measurement device fails. On the IEEE 15 bus feeder, this paper compares TANFL with a conventional multilayer perceptron (MLP) and the formula-based localization method under identical data, measurement availability scenarios, and evaluation metrics. On the test partition, TANFL outperforms the conventional MLP; with complete measurements, its line fault location performance is slightly lower than that of the formula-based method, but in the average results for scenarios with one missing bus measurement block, its accuracy is the highest among the three methods at a 5% tolerance.

I. INTRODUCTION

Distribution-feeder fault location is complicated by branching networks, variable conditions, and potentially unavailable measurements. The paper identifies a gap in combining continuous location, bus-versus-line identification, and missing-input robustness in one locator, and proposes TANFL to address it.

I. INTRODUCTION

  • Branching feeders, variable line and load conditions, unknown fault resistance, and unavailable devices complicate fault localization and can leave multiple locations electrically similar.The result must identify the affected component as well as its location.Source: §I · p. 1
  • Prior studies cover continuous location, bus-versus-line identification, and missing-input robustness in different combinations, but none combines all three in one model.The reviewed learning methods include examples that omit busbar faults or continuous position output under missing-channel conditions.Source: §I · p. 1
  • TANFL uses post-transient voltage and current phasors, feeder connections, and line parameters to combine component identification, continuous line-position estimation, and missing-input handling.Its stated contributions include missing indicators and random bus-block masking, with comparisons against a conventional MLP and an adapted formula-based locator.Source: §I · p. 2

II. TOPOLOGY-AWARE NEURAL NETWORK-BASED FAULT LOCATOR

The TANFL architecture uses a per-bus input vector built from preprocessed phasor features and a missing-data indicator. These vectors are stacked across buses for network input.

II. TOPOLOGY-AWARE NEURAL NETWORK-BASED FAULT LOCATOR

2) Shared Bus Encoder:

TANFL shares a bus encoder across feeder nodes, then uses separate bus and line readouts. The gated line readout incorporates topology and line parameters to score lines and estimate positions.

2) Shared Bus Encoder:

  • The shared dense bus encoder maps each bus input to an embedding, with shared weights keeping it independent of bus count.The bus readout combines each bus embedding with the feeder mean embedding to produce a bus score.Source: 2) Shared Bus Encoder · p. 2, 3) Bus Readout · p. 2

3) Bus Readout:

  • The bus readout concatenates each bus embedding with the feeder mean embedding and outputs one score per bus.The feeder mean is the mean of the bus embeddings.Source: 3) Bus Readout · p. 2

4) Gated Line Readout:

  • The gated line readout combines endpoint and child-bus embeddings, feeder context, and line resistance/reactance to predict a line score and position.A projected endpoint-embedding difference and a learned gate are used with topology features; the gate input also includes node and log-magnitude features.Source: 4) Gated Line Readout · p. 2

5) Joint Classification:

TANFL jointly scores bus and line fault components, selects the highest-probability component, and reports a clipped position estimate when that component is a line.

5) Joint Classification:

  • A softmax over bus and line scores gives the probability of each faulted component, and the highest-probability component is selected.The output includes the component probability vector and one position output per line.Source: 5) Joint Classification · p. 2
  • For a selected line, TANFL takes that line’s position output and clips it to [0, 1].The endpoints 0 and 1 denote the upstream and downstream ends, respectively.Source: 5) Joint Classification · p. 2
  • Direct bus faults have their own classes and do not require a position estimate.The network output contains component probabilities and unclipped positions for all lines.Source: 5) Joint Classification · p. 2

B. Data Generation and Preprocessing

The study simulates single-line-to-ground faults on lines and buses and extracts post-transient voltage and current phasors. Phasors are phase-referenced and scaled using constants fitted on the training partition.

B. Data Generation and Preprocessing

  • Fault cases cover single-line-to-ground faults on lines and busbars, with fault locations spanning each line.Post-transient phasors are extracted from recorded AC fundamental components after the fault-induced DC component sufficiently decays.Source: B. Data Generation and Preprocessing · p. 3
  • Voltage and incoming-current channels are measured at each bus, and the phase A voltage of the supply bus sets the phase reference.The recorded phasors are rotated to that reference before scaling.Source: B. Data Generation and Preprocessing · p. 3
  • Voltage components are centered and scaled, while centered current components are compressed with asinh before scaling.Scaling constants are fitted on the training partition.Source: B. Data Generation and Preprocessing · p. 3

C. Training and Validation Procedure

Preprocessing constants are fitted only after the data are partitioned, using training data alone. Training combines classification, feeder-path, and line-position losses and exposes the network to complete and masked inputs.

C. Training and Validation Procedure

  • Preprocessing constants are fitted after partitioning so scaling statistics come only from the training partition.The dataset is divided into training, validation, and test partitions first.Source: C. Training and Validation Procedure · p. 3
  • The loss combines categorical classification, probability-weighted normalized path distance, and absolute position error on the true faulted line.Path distance sums series-impedance magnitudes along the feeder route between predicted and true locations.Source: C. Training and Validation Procedure · p. 3
  • Each training sample has a 0.5 probability of masking a bus measurement where missing values are permitted, balancing complete and incomplete inputs.Validation path MAE is calculated for complete inputs and scenarios with one missing bus measurement block; the test partition is evaluated with complete measurements and sequential bus-block removals.Source: C. Training and Validation Procedure · p. 3

III. CASE STUDY

The case study evaluates TANFL against an MLP and a formula-based locator on an IEEE 15-bus feeder. Simulated cases span line and bus faults, with tests including removal of one bus’s measurement block.

III. CASE STUDY

  • The case study compares trained TANFL with a conventional MLP and formula-based locator under complete measurements and one-block-missing scenarios.The comparison is conducted on the test partition.Source: §III · p. 3

A. Test System

  • The IEEE 15-bus feeder has 15 buses, 14 lines, and a rated voltage of 13.8 kV; buses 1–4 remain measured in every missing-data scenario.Each bus provides three-phase voltage and incoming-current phasors plus a missing indicator; one complete voltage-and-current block is removed from buses 5–15 in missing-data tests.Source: A. Test System · p. 3

1) Dataset Acquisition:

  • The 1,840 simulated cases comprise 110 line-fault cases per line and 20 direct bus-fault cases per bus.Only phase-A-to-ground faults are simulated; fault resistance ranges from 0.01–20 Ω and load factors from 80%–120%. Cases are split approximately 70/15/15 across training, validation, and test, with every bus represented in each partition.Source: 1) Dataset Acquisition · p. 3

2) Training settings:

TANFL is trained with a compact parameterization and compared with an MLP and a feeder-tailored formula-based locator. Evaluation uses tolerance-based fault accuracy and electrical path MAE.

2) Training settings:

  • TANFL uses an embedding dimension of 32 and has 13,795 trainable parameters.Training uses Adam with an initial learning rate of 10^-3, batch size 64, and up to 300 epochs; the checkpoint with lowest validation loss is saved.Source: 2) Training settings · p. 3

3) Comparison of localization methods:

  • The model contains 128 and 64 swish units, respectively, and has 47,659 parameters, 3.5 times TANFL’s, while being trained with loss function (9) and the same rules.Formula-based position estimation requires measurements at both ends of the line; candidates lacking necessary measurements are excluded. Its busbar and line selection rules are tailored to this feeder.Source: 3) Comparison of localization methods · p. 4

4) Evaluation metrics:

  • Line-fault accuracy requires the predicted line to match and position error to fall within tolerance; bus-fault hits also allow a nearby connected-line prediction.Accuracy is reported at 1%, 3%, 5%, and 10% tolerances, while path MAE averages electrical path distance, with smaller values indicating shorter average distance.Source: 4) Evaluation metrics · p. 4

B. Localization with Complete Measurements

With complete measurements, TANFL performs comparably to the formula-based locator at 5% and 10% tolerances, and both outperform the conventional MLP for line and bus faults. At 5%, TANFL is close to the formula-based locator for line accuracy and matches it for bus accuracy.

B. Localization with Complete Measurements

  • At 5% and 10% tolerances, TANFL is comparable to the formula-based locator, and both outperform the MLP for line and bus faults.The figure reports test-partition accuracy across 1%, 3%, 5%, and 10% tolerances for line and bus faults.Source: B. Localization with Complete… · p. 4, Figure 3 · p. 4
  • Most TANFL misses at tight tolerances are position errors on the correctly identified line, so accuracy rises as tolerance relaxes.The formula-based locator’s remaining line-fault misses occur near the line start, where its 2% decision threshold classifies them as upstream bus faults.Source: B. Localization with Complete… · p. 4
  • At 5% tolerance, TANFL reaches 97.48% line and 100% bus accuracy, versus the MLP’s 80.67% and 93.33%; formula-based line accuracy is 97.90%.Both TANFL and the formula-based locator achieve 100% bus accuracy.Source: B. Localization with Complete… · p. 4, Table I
  • Path MAE is 1.37 × 10^-3 p.u. for TANFL, below the MLP’s 3.21 × 10^-3 p.u.; the formula-based locator achieves 0.05 × 10^-3 p.u.The formula-based result is reported under the tested simulation conditions.Source: B. Localization with Complete… · p. 4

C. Localization with Missing Measurements

With one missing bus measurement block, TANFL leads the other methods on average for most tested fault and tolerance combinations. The reported results average eleven single-block-missing scenarios, while the formula-based locator can evaluate the true faulted element in only some cases.

C. Localization with Missing Measurements

MISSING BUS MEASUREMENT BLOCK

  • At 5% tolerance, TANFL reaches 88.35% line and 92.73% bus accuracy, above the MLP’s 76.89% and 89.90% and the formula-based locator’s 79.49% and 86.67%.TANFL’s path MAE is 2.80 × 10^-3 p.u., versus 4.33 × 10^-3 p.u. for the MLP and 49.03 × 10^-3 p.u. for the formula-based locator.Source: Table II · p. 4, MISSING BUS MEASUREMENT BLOCK · p. 5
  • The true faulted element is computable in only 82.04% of cases for the formula-based locator.This percentage describes computational availability rather than accuracy.Source: MISSING BUS MEASUREMENT BLOCK · pp. 4–5
  • TANFL’s 5% line-fault accuracy is lowest when Bus 6 is missing, at 79.83%.The formula-based locator requires specific currents for its current-balance residual and both terminal voltages for line-position estimation, so missing measurements exclude candidate elements.Source: MISSING BUS MEASUREMENT BLOCK · p. 5
  • Across eleven missing-block scenarios, TANFL leads for line faults at 3%, 5%, and 10% tolerances and for bus faults at every tested tolerance.The formula-based locator leads only for line faults at 1%; with complete measurements, it leads for line faults at 1%, 3%, and 5%, while TANFL leads at 10%.Source: MISSING BUS MEASUREMENT BLOCK · p. 5

IV. CONCLUSION

The conclusion presents TANFL as a post-transient-phasor fault locator that identifies a faulted bus or line and estimates position along the selected line, including with one missing measurement block. On the IEEE 15-bus feeder, it outperforms the MLP at 5% tolerance in complete and missing-measurement tests.

IV. CONCLUSION

  • On the IEEE 15-bus feeder, TANFL exceeds the conventional MLP at 5% tolerance with fewer parameters under both complete and missing measurements.With complete measurements, its line-fault accuracy is comparable to the formula-based locator at 5% and 10%; across eleven missing-block scenarios, it leads line accuracy at 3%, 5%, and 10% and bus accuracy at all tested tolerances.Source: §IV · p. 5
  • Future work will extend TANFL to other fault types.The conclusion’s reported findings are for the tested IEEE 15-bus feeder and measurement conditions.Source: §IV · p. 5
Loading 2610.36474v1…