PaperlayerView exact revision on arXiv

Source-linked AI summary

GCN-Based Model for Fault Classification in Distribution Networks under Harmonic Distortion

Fathima Razeeya Mohamed Razick, Syed Muhammad Ahsan, Petr Musilek

arXiv:2609.27169v1eess.SP

TL;DR

Distribution faults are difficult to classify as inverter-based PV changes waveforms and introduces harmonic distortion, while prior work largely addressed ideal or mildly distorted conditions. The paper maps phasor features to a topology-aware GCN and evaluates it across PV and THDi levels. The model achieves about 0.98 F1-score and outperforms CNN, LSTM, and MLP models in the most stressed case.

  • Problem

    Inverter-based PV harmonics and dynamic variations alter voltage and current waveforms, complicating fault classification, while prior studies largely lacked evaluation under these conditions.

    Source: §I · pp. 1–2

  • Method

    The paper maps pre-fault and post-fault voltage and current phasors, including harmonics, onto the IEEE 34-bus topology and uses differential evolution to select GCN design parameters.

    Source: B. Design Parameters · p. 3, §V · p. 4

  • Results

    About 0.98 F1-score was achieved, and the GCN outperformed CNN, LSTM, and MLP models in the 70% PV, 5% THDi case.

    Source: §IV · p. 4, §V · p. 5

  • Takeaways & Limitations

    F1-scores remained above 0.97 across the reported cases, with confusion limited to closely related classes at the highest stress settings.

    Source: §IV · p. 4, §V · p. 5

Abstract

from arXiv · show

The increasing integration of distributed energy resources introduces new challenges for accurate fault detection and classification in distribution networks. This paper presents a topology-aware graph convolutional network-based model for fault classification under inverter-induced harmonic distortion. The model uses pre-fault and post-fault voltage and current phasors, including harmonic components, mapped to the IEEE 34-bus system topology. The simulation data generated in OpenDSS encompass varying levels of penetration of solar photovoltaics (30%, 50%, 70%) and current total harmonic distortion (0%, 1%, 3%, 5%) across multiple fault types: single line-to-ground, line-to-line, double line-to-ground, and three-phase-to-ground. A differential evolution algorithm is employed to optimize the hyperparameters of the network, resulting in a three-layer model that achieves an F1-score of 0.98. Stress testing across scenarios confirms high robustness and minimal class confusion. The results demonstrate that a graph convolutional network-based model can effectively classify distribution faults under varying photovoltaic penetration levels and harmonically distorted conditions. Comparative analyses with convolutional neural network, long short-term memory, and multi-layer perceptron architectures demonstrate that the proposed model delivers superior classification performance, particularly under the most stressed condition with 70% photovoltaic penetration and 5% harmonic distortion.

I. INTRODUCTION

Distribution faults are a major source of power interruptions, and growing PV integration makes their signatures harder to distinguish. The paper addresses gaps in prior fault-classification work by evaluating a topology-aware GCN under jointly varying PV penetration and harmonic distortion.

I. INTRODUCTION

  • Approximately 80% of power supply interruptions across the power system are caused by distribution grid faults, which can degrade power quality, damage equipment, and trigger large-scale outages.Faults include SLG, LL, LLG, and LLLG categories.Source: §I · p. 1
  • Inverter-based PV harmonics and dynamic variations alter voltage and current waveforms, making fault signatures harder to distinguish and limiting methods based on fundamental-frequency quantities.The paper motivates intelligent, data-driven, topology-aware classification for these conditions.Source: §I · p. 1
  • Prior studies largely considered ideal or mildly distorted conditions, leaving inverter-induced current harmonics and PV-generation dynamics insufficiently evaluated in fault classification.The paper evaluates a phasor-based GCN jointly across PV penetration and THDi levels, with comparisons to CNN, LSTM, and MLP models.Source: §I · p. 2

II. GCN-BASED MODEL ARCHITECTURE AND DESIGN PARAMETERS

The model uses graph structure to combine bus-level electrical features with information from connected buses. Self-loops and symmetric normalization support message passing across the feeder.

II. GCN-BASED MODEL ARCHITECTURE AND DESIGN PARAMETERS

  • GCNs extend convolution to irregular graphs, representing both each bus’s properties and neighboring-bus effects such as voltage, current, and harmonic components.Distribution buses are nodes and lines are edges in this representation.Source: §II · p. 2
  • Stacked graph-convolution layers capture multi-hop relationships and summarize local electrical behavior with broader grid interactions for fault classification.This representation is described as capturing disturbance propagation across connected buses.Source: §II · p. 2
  • Self-loops let each bus retain its own information during message passing, while symmetric normalization balances the effects of highly connected nodes.The representation combines local electrical quantities with spatial disturbance propagation.Source: §II · p. 2

A. Proposed GCN-Based Model Architecture

The proposed GCN-based model is organized into three stages to jointly capture spatial relationships among buses and electrical behavior under fault and harmonic conditions.

A. Proposed GCN-Based Model Architecture

1) Graph-Based message passing:

Graph message passing aggregates information according to feeder topology, and stacking layers expands the model’s spatial receptive field. This lets the network represent fault-related electrical perturbations across multiple feeder hops.

1) Graph-Based message passing:

  • Each graph-convolution layer exchanges information between directly connected neighbors using the feeder’s electrical topology.The update uses normalized adjacency, node features, trainable weights, a bias, and a nonlinear activation.Source: 1) Graph-Based message passing · p. 2

2) Hierarchical feature aggregation:

  • With three layers, the receptive field extends up to three electrical hops, capturing voltage and current perturbations spreading from a faulted bus.Multi-hop aggregation can help distinguish locally similar LL and LLG patterns through their different regional signatures.Source: 2) Hierarchical feature aggregation · p. 2

3) Graph-level classification:

Differential evolution selects a three-layer GCN that achieves an F1-score of 0.98. Mean pooling combines bus embeddings into a grid-level representation for fault classification.

3) Graph-level classification:

  • Mean pooling combines enriched bus embeddings into a single representation of the system state for classification.A fully connected layer maps the pooled representation to class logits, and argmax selects the predicted fault class.Source: 3) Graph-level classification · p. 2

B. Design Parameters

  • The differential-evolution-selected three-layer GCN achieves a test-set F1-score of 0.98.The model uses 256 filters, 0.41 dropout, a 1.4·10−3 learning rate, 7.8·10−3 weight decay, and batch size 32.Source: B. Design Parameters · p. 3
  • The proposed architecture stacks three graph convolutional layers before global pooling and fully connected layers.The architecture is illustrated in Fig. 1.Source: B. Design Parameters · p. 3, Figure 1 · p. 3

III. SIMULATION SETUP

The study evaluates fault detection and classification on the standard IEEE 34-bus feeder. The system model is implemented in OpenDSS for steady-state and fault simulations.

III. SIMULATION SETUP

  • The simulation setup covers test-system configuration, simulation cases, and preparation of training data.This section introduces the study’s overall simulation setup.Source: §III · p. 3

A. Test system Configuration

  • The test system is the standard 24.9 kV medium-voltage unbalanced IEEE 34-bus feeder, with single- and three-phase nodes and lines, transformers, and multiple load types.Its single-line diagram is shown in Fig. 2.Source: A. Test system Configuration · p. 3, Figure 2 · p. 3
  • The feeder model is implemented in OpenDSS, which supports steady-state and fault simulations through the OpenDSSDirect Python interface.Source: A. Test system Configuration · p. 3

B. Simulation Cases

The simulation cases vary PV penetration and current total harmonic distortion on the IEEE feeder. The resulting dataset spans specified fault conditions and generated samples across distortion levels.

B. Simulation Cases

  • The cases span 30%, 50%, and 70% PV penetration and 0%, 1%, 3%, and 5% THDi.Each 200 kWp PV installation is assigned to randomly selected buses; penetration is measured against the 1769 kW system peak load. Fig. 3 shows spectra for the 1%, 3%, and 5% distortion levels.Source: B. Simulation Cases · p. 3, Figure 3 · p. 3

C. Dataset Preparation

  • OpenDSS simulations represent normal operation plus 10 fault configurations—three SLG, three LL, three LLG, and LLLG—across valid buses; fault resistance is randomly chosen from [0.05, 25] Ω.Source: C. Dataset Preparation · p. 3
  • The dataset contains 11,000 cases, with 250 fault samples per fault type at each THDi level.Source: C. Dataset Preparation · p. 4

D. Feature Extraction and Preprocessing

Pre-fault and post-fault electrical measurements, including harmonic components, are transformed into graph features. The feature matrix is linked to feeder topology and fault labels for GCN learning.

  • Each simulation contributes pre-fault and post-fault voltage and current phasors, including harmonic components, at node and line levels.Source: D. Feature Extraction and Preprocessing · p. 4
  • Per-bus features include phase-wise voltage magnitude and angle differences, line-current magnitude and angle differences, and harmonic currents for orders 3, 5, 7, 9, 11, and 13.Harmonic-current features include pre-fault, post-fault, and difference values when harmonic fields exist.Source: D. Feature Extraction and Preprocessing · p. 4
  • The standardized feature matrix is mapped to OpenDSS-derived feeder edges and fault labels, then split 80%/10%/10% for training, validation, and testing.Validation and test data use the training-set mean and standard deviation for standardization.Source: D. Feature Extraction and Preprocessing · p. 4

IV. RESULTS AND DISCUSSION

The proposed GCN converges rapidly and maintains high F1-scores across PV-integrated and non-PV operating conditions. Under the most stressed case, it outperforms CNN, LSTM, and MLP models for normal and most fault conditions, with remaining errors concentrated among similar fault classes.

IV. RESULTS AND DISCUSSION

  • Training and validation converge rapidly within the first ten epochs in both non-PV and PV-integrated scenarios, with only slight changes in convergence rate under PV integration.The reported loss curves show minor oscillations and lower loss values, indicating learning without overfitting.Source: §IV · p. 4, Figure 4 · p. 4
  • Average F1-scores remain above 0.97 across all tested cases, although higher THDi slightly reduces them.The test evaluates 550 simulated fault scenarios for each PV-penetration and THDi combination.Source: §IV · p. 4
  • Confusion matrices show diagonal dominance in non-PV and most-stressed conditions, with minor errors mainly between closely related LL and LLG faults.The most-stressed condition is 70% PV with 5% THDi.Source: §IV · p. 4
  • Under 70% PV and 5% THDi, the proposed GCN outperforms CNN, LSTM, and MLP models for normal and most fault conditions.The comparison uses per-class F1-scores.Source: §IV · p. 4

V. CONCLUSIONS

The topology-aware GCN classifies faults from pre- and post-fault phasors, including harmonics, mapped to the IEEE 34-bus network. Its three-layer configuration achieves an F1-score of about 0.98, and comparative evaluation reports superior accuracy and generalization against CNN, LSTM, and MLP models.

V. CONCLUSIONS

  • The model classifies distribution faults using pre- and post-fault voltage and current phasors, including harmonic components, mapped to the IEEE 34-bus network.The evaluated fault categories are SLG, LL, LLG, and three-phase-to-ground.Source: §V · pp. 4–5
  • An F1-score of about 0.98 was achieved by the selected three-layer model, and comparison confirmed superior accuracy and generalization against CNN, LSTM, and MLP models.Differential evolution selected the model configuration.Source: §V · p. 5
  • Performance remains stable across 30%, 50%, and 70% PV penetration and 0%, 1%, 3%, and 5% current THD, with confusion limited to closely related classes at the highest stress settings.Table I summarizes accuracy and F1-score under varying PV penetration and THDi levels.Source: §V · p. 5, Table I · p. 5
Loading 2609.27169v1…