Pith. sign in

REVIEW 2 major objections 5 minor 29 references

Jointly modeling grid topology and time keeps fault location accurate when distributed-energy penetration shifts, while models using only one modality lose accuracy sharply.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 09:48 UTC pith:QYSUGMOI

load-bearing objection Solid benchmark, but the asymmetric generalization result is partly an artifact of the nested DER placement design. the 2 major comments →

arxiv 2607.29293 v1 pith:QYSUGMOI submitted 2026-07-31 cs.LG cs.SYeess.SY

Assessing the Generalization of Graph Neural Networks for Fault Location Across Increasing Distributed Energy Resource Penetration Levels

Burak Karabulut , Olayiwola Arowolo , Carlo Manna , Chris Develder , Jochen L. Cremer This is my paper
classification cs.LG cs.SYeess.SY
keywords fault locationdistribution networksdistributed energy resourcesgraph neural networksspatio-temporal modelingasymmetric generalizationmeasurement noiseDER penetration
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that joint spatio-temporal modeling is what keeps neural fault-location models reliable as distributed energy resources (DERs) reshape fault signatures. It compares a spatio-temporal graph attention network (STGATv2) with a purely temporal GRU, a purely spatial GATv2, and classical machine-learning baselines on a reconfigured IEEE 123-bus feeder with 10%, 25%, and 50% DER penetration. In-distribution, STGATv2 reaches 92-94% macro F1. The key finding is asymmetric generalization: models trained at 50% penetration retain near in-distribution performance at lower penetrations, while models trained at 10% degrade sharply at 50%; STGATv2 still keeps 81-84% F1 there, and under measurement noise keeps above 85% F1 while the temporal GRU collapses to as low as 33.5% F1. The takeaway is that topological awareness plus temporal memory is the decisive ingredient for fault location in active distribution networks.

Core claim

The central claim is that STGATv2, which first extracts temporal embeddings from voltage time series with a GRU and then propagates them through a graph attention layer over the measured buses, outperforms purely spatial, purely temporal, and traditional ML baselines for fault location, and does so more reliably under distribution shift in DER penetration and under measurement noise. The asymmetry result, where training at high penetration generalizes downward but training at low penetration does not generalize upward, is explained by high-penetration configurations containing the fault patterns found at lower levels. The paper argues that this establishes joint spatio-temporal modeling, not

What carries the argument

STGATv2 couples a gated recurrent unit (GRU), which produces a latent temporal representation per measurement node, with a GATv2 graph attention layer, which weights neighbor contributions based on current node features. The graph is built only from the 25 measurement locations (the 'measured-only' strategy), reflecting partial observability, and node-level class probabilities are aggregated by soft voting. The GRU supplies temporal memory for fault dynamics; the attention layer supplies topological awareness by letting fault signatures propagate across the feeder. Comparisons to the GRU alone and GATv2 alone isolate the contribution of each component.

Load-bearing premise

The inference that high-penetration training generalizes 'downward' rests on the data-construction choice that lower DER penetration configurations are subsets of higher ones; if real DER deployments do not nest, the 50% model is being tested on spatial layouts already seen in training, so the apparent downward generalization may be interpolation, not a model property.

What would settle it

Repeat the 10%-to-50% generalization experiment with a 50% DER hosting set that shares no nodes with the 10% set while keeping aggregate penetration and fault types fixed. If STGATv2's F1 under that shift stays close to the nested-case 81-84% and still clearly beats GATv2/GRU, the paper's topological-awareness conclusion is supported; if its F1 drops to baseline levels, the asymmetry result is an artifact of nested DER placement.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the asymmetric generalization holds, utilities can train fault-location models at high DER penetration and deploy them as penetration rises, with only a small loss in accuracy.
  • Models used in active distribution networks should include graph topology; purely temporal models risk severe degradation under noise, down to 33.5% F1, even when clean-data performance looks acceptable.
  • DER placement matters: dispersed configurations generalize more easily across penetration shifts than localized clusters, so hosting-node planning can double as a data-quality lever.
  • The 10%-to-50% shift is the hardest direction; evaluating fault-location models only in-distribution overstates their readiness for grids undergoing DER growth.
  • Soft voting across measured nodes reduces outlier influence, suggesting that distributed monitoring with aggregation is a useful architectural choice under partial observability.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the low-penetration DER sets are nested inside the high-penetration set, the downward-generalization result may partly reflect interpolation within the training support; a test with non-nested DER deployments would clarify whether the effect is a property of the model or of the data construction.
  • The same asymmetry likely applies to other distribution-grid tasks that depend on DER deployment, such as voltage estimation or state inference: models trained on rich, high-penetration scenarios may transfer down more easily than models trained on sparse scenarios transfer up.
  • A concrete extension is to train STGATv2 at 50% penetration and evaluate at 100% penetration, as the authors suggest, to see whether the monotonic containment still holds and whether the F1 drop accelerates once DER locations move beyond the training support.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. This paper empirically benchmarks a spatio-temporal graph attention network (STGATv2) against purely temporal (GRU), purely spatial (GATv2), and traditional ML baselines for fault location on a modified IEEE 123-bus distribution feeder with DER penetration levels of 10%, 25%, and 50%. Synthetic fault data are generated with OpenDSS/PyDSS, and macro F1 is reported over 3 random seeds for two DER placement configurations. The paper claims that STGATv2 consistently outperforms the baselines in-distribution (92-94% F1), that cross-penetration generalization is asymmetric (training at 50% retains performance at lower levels, while training at 10% degrades at 50%), and that STGATv2 is substantially more robust to measurement noise than the baselines.

Significance. If the results are robust, the paper is a useful empirical contribution: it extends STGNN-based fault location to varying DER penetration, considers two DER placement configurations, and provides a relatively detailed synthetic data-generation pipeline. The in-distribution architecture comparison is informative and the use of macro F1 with standard deviations in Table II is a strength. However, the headline claim of generalization to 'unseen' penetration levels is confounded by the nested DER hosting-node construction described in Section III-A, so the paper's second contribution is not yet established. The noise experiment also lacks the statistical detail needed to support its dramatic claims.

major comments (2)
  1. [Section III-A / IV-B / Abstract] The core claim of generalization to 'unseen' DER penetration levels is confounded by the data-generation design. Section III-A states that 'higher generation is achieved by progressively expanding the hosting node set, such that lower level configurations are subsets of higher ones.' Table I confirms that the 10% and 25% DER hosting nodes are contained in the 50% node set. Thus a model trained at 50% has already seen fault signatures at the exact spatial locations that appear at 10% and 25%, including zero-output cases; the low-penetration test conditions are within the spatial support of the training distribution even if the exact combination of active DERs differs. The observation in Section IV-B that 'high penetration ... includes fault patterns seen at lower levels' is therefore a direct artifact of the nested construction, not an empirical discovery about deployment-realistic DER ro
  2. [Section IV-C] The noise-robustness claim rests on point estimates with no error bars, no table, and no statement of the number of seeds or the exact train/test penetration level and DER configuration used. The text reports specific values (e.g., 'GRU drops as low as 33.5% F1') but no standard deviations, making it impossible to assess whether the dramatic GRU collapse is reproducible. Please report mean±std over the same 3 seeds (or more), specify the SNR injection protocol and the penetration/configurations used, and ideally provide a table analogous to Table II so that the noise experiment is on the same evidentiary footing as the main results.
minor comments (5)
  1. [Section II-A] The construction of the 'measured only' graph is ambiguous: if the node set V consists only of the 25 measurement locations, the adjacency definition 'A_{u,v}=1 if buses u and v are directly connected' needs clarification for measurement nodes that are not adjacent in the original feeder. Please specify how edges are formed for non-adjacent measurement nodes.
  2. [Section IV-A] The RF and PCA-SVM results are reported in text only, without standard deviations or a table. Since these baselines are used to motivate the neural comparison, a small supplementary table with error bars would improve reproducibility.
  3. [Section II-C] In Eq. (2), the activation function φ is not specified. Please state whether it is ReLU or another nonlinearity, and whether it is applied after the attention aggregation.
  4. [Section IV-C] Please specify which trained model (which penetration level and which configuration) is used for the noise experiments. The phrase 'maintaining baseline DER penetration levels' is vague.
  5. [References] References [21] and [22] list publication years 2027 and 2026, respectively; if these are forthcoming or already published, please verify the citation details and add DOIs where available.

Circularity Check

0 steps flagged

No circular derivation: the paper reports measured held-out F1 scores, and the only self-citation is a non-load-bearing graph-construction choice.

full rationale

The paper's central claims are empirical benchmark results (e.g., STGATv2 92–94% F1 in-distribution; generalization drops under cross-penetration tests). Each number is obtained by training on one simulated penetration level and evaluating on a separately simulated held-out test set; no fitted parameter is renamed as a prediction and no test-set quantity is used to define the models. The only same-author citation is [24], for the 'measured only' graph strategy. That strategy is an architectural input applied uniformly to the graph-based models, not a premise that the paper's experiments recover or that forces the STGATv2-vs-baseline ordering; thus it is not load-bearing. The nested DER placement ('lower level configurations are subsets of higher ones', Section III-A) means the 50%-trained model has seen the low-penetration hosting nodes, so the downward-generalization result may be optimistic; this is a threat to external validity, not a circular reduction: the test data are still independently simulated with different nominal capacities, M-scaling, loads, and fault resistances, and the F1 is measured rather than derived. No equation in Section II reduces to its input, and no uniqueness theorem or ansatz is imported from the authors' prior work. Therefore the derivation chain is self-contained and no circular step is exhibited.

Axiom & Free-Parameter Ledger

11 free parameters · 6 axioms · 0 invented entities

The central claims rest on a chain of simulation and modeling choices. The free parameters listed are hand-chosen experimental settings (not fitted constants); they determine the data distribution and thus the reported F1 scores. The axioms are domain assumptions about realism and representativeness that are not validated against real-world fault data.

free parameters (11)
  • learning_rate = 0.0005
    AdamW learning rate chosen empirically (Section III-B); affects convergence and final F1.
  • weight_decay = 1e-4
    AdamW weight decay, chosen empirically (Section III-B).
  • dropout = 0.35
    Dropout rate for all neural models, chosen empirically (Section III-B).
  • attention_dropout_and_heads = 0.3, 4 heads
    GATv2 attention dropout and head count (Section III-B).
  • GRU_hidden_size = 128
    Recurrent hidden size for STGNN and GRU baseline (Section III-B).
  • GNN_hidden_size = 64/128
    GNN layer hidden dimension (Section III-B).
  • load_scaling_range = U(0.5, 1.3)
    Simulation parameter for load variation (Section III-A); affects fault signature diversity.
  • DER_output_scaling_M = {0, 0.25, 0.5, 0.75, 1}
    Shared DER output scaling factor (Section III-A); defines penetration-level operating conditions.
  • fault_resistance_set = {10,20,40,80,100} Ω
    Fault impedance levels (Section III-A); moderate-to-high impedance is central to the difficulty of the task.
  • BESS_probabilities = 0.6/0.3/0.1 and 0.3/0.6/0.1
    Conditional charging/discharging/idle probabilities (Section III-A); introduce stochasticity.
  • noise_SNR_levels = 50 dB, 45 dB
    Test-time Gaussian noise levels (Section IV-C); the robustness claim depends on these specific levels.
axioms (6)
  • domain assumption OpenDSS/PyDSS dynamic simulations faithfully represent real distribution network fault behavior
    All data is synthetic; no real-world validation is provided (Section III-A).
  • domain assumption The reconfigured IEEE 123-bus feeder with the specified reconfiguration is a representative testbed for fault location difficulty
    The switch changes (open 60-160, close 54-94) are claimed to produce more subtle voltage drops, but this is a single topology (Section III-A).
  • domain assumption Macro F1 on the synthetic test set is the correct measure of fault location accuracy
    The paper uses macro F1 throughout; no error analysis in terms of physical distance to fault is provided (Section III-B).
  • domain assumption The 'measured only' graph strategy from [24] is valid for partial observability
    The graph uses only measurement nodes, relying on the authors' prior work; no ablation of full topology graph is given (Section II-A).
  • domain assumption Nested DER placement (lower configurations are subsets of higher ones) is representative of real DER deployment
    This design choice is key to the asymmetry interpretation (Section III-A).
  • domain assumption Soft voting across node-level predictions is an appropriate aggregation
    No comparison with other aggregation schemes (e.g., hard voting or max-pooling) is reported (Section II-A).

pith-pipeline@v1.3.0-daily-deepseek · 9727 in / 19126 out tokens · 189247 ms · 2026-08-03T09:48:59.982435+00:00 · methodology

0 comments
read the original abstract

Accurate fault location is critical for distribution network reliability. However, increasing distributed energy resource (DER) penetration complicates fault location due to intermittent generation and bidirectional power flows that reshape fault signatures. Spatio-Temporal Graph Neural Networks (STGNNs) have shown promise by jointly modeling spatial and temporal dependencies, but their behavior under increasing DER penetration has not been studied rigorously. In this paper, we (i) systematically benchmark spatio-temporal graph attention network (STGATv2) against purely temporal (gated recurrent unit, GRU), purely spatial (GATv2) and traditional machine learning baselines, and (ii) evaluate how well models generalize across increasing DER penetration levels (10%, 25%, 50%) on a reconfigured IEEE 123-bus feeder with multiple DER injection points and moderate-to-high impedance faults. Results show that STGATv2 consistently outperforms neural baselines, achieving 92-94% macro F1 in-distribution. Notably, generalization across penetration levels is asymmetric: training at 50% penetration retains near in-distribution F1 score at lower levels, whereas training at 10% degrades considerably at 50% - with STGATv2 retaining 81-84% F1 under these drastic shifts, substantially higher than GATv2 and GRU which drop to 69-74% F1 and 73-75% F1 respectively. Under realistic measurement noise, STGATv2 maintains > 85% F1, while GRU drops as low as 33.5% F1, highlighting the critical role of topological awareness for robust fault location in active distribution networks.

Figures

Figures reproduced from arXiv: 2607.29293 by Burak Karabulut, Carlo Manna, Chris Develder, Jochen L. Cremer, Olayiwola Arowolo.

Figure 1
Figure 1. Figure 1: Model architectures for fault location: (a) Shared GRU for temporal feature extraction per node; (b) Shared GATv2, where measurement sequences [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: IEEE 123-node feeder with fault, DER, and measurement locations, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 2 linked inside Pith

  1. [1]

    IEEE Std 1366-2022 (Revision of IEEE Std 1366-2012)

    IEEE, Piscataway, NJ, USA,IEEE Guide for Electric Power Distribution Reliability Indices, 2022. IEEE Std 1366-2022 (Revision of IEEE Std 1366-2012)

    2022

  2. [2]

    J. J. Grainger and W. D. Stevenson,Power System Analysis. New York: McGraw-Hill, 1994

    1994

  3. [3]

    Electricity grids and secure energy tran- sitions,

    International Energy Agency, “Electricity grids and secure energy tran- sitions,” tech. rep., IEA, Paris, France, 2023

    2023

  4. [4]

    Graphgan-based fault detection and location for complex power grids,

    J. Liu, H. Shi, Y . Chen, C. Yang, M. Ma, and Y . Li, “Graphgan-based fault detection and location for complex power grids,” inProc. Int. Conf. Energy Power Electr . Technol. (CEPET 2025), (Wuhan, China), pp. 465– 469, IEEE, 2025

    2025

  5. [5]

    Impact assessment of increasing renewable energy penetration on voltage insta- bility tendencies of power system buses using a QV-based index,

    B. B. Adetokun, C. M. Muriithi, J. O. Ojo, and O. Oghorada, “Impact assessment of increasing renewable energy penetration on voltage insta- bility tendencies of power system buses using a QV-based index,”Sci. Rep., vol. 13, p. 9782, 2023

    2023

  6. [6]

    Review on artificial intelligence-based fault location methods in power distribution net- works,

    H. Rezapour, S. Jamali, and A. Bahmanyar, “Review on artificial intelligence-based fault location methods in power distribution net- works,”Energies, vol. 16, no. 12, p. 4636, 2023

    2023

  7. [7]

    An extended impedance- based fault location algorithm in power distribution system with dis- tributed generation using synchrophasors,

    S. Chandran, R. Gokaraju, and K. Narendra, “An extended impedance- based fault location algorithm in power distribution system with dis- tributed generation using synchrophasors,”IET Gener . Transm. Distrib., vol. 18, no. 3, pp. 479–490, 2024

    2024

  8. [8]

    A new method for fault location in distribution networks based on voltage sag measurements,

    R. F. Buzo, H. M. Barradas, and F. B. Le ˜ao, “A new method for fault location in distribution networks based on voltage sag measurements,” IEEE Trans. Power Deliv., vol. 36, no. 2, pp. 651–662, 2021

    2021

  9. [9]

    A novel fault location method based on traveling wave for multi-branch distribution network,

    F. Liu, L. Xie, K. Yu, Y . Wang, X. Zeng, L. Bi, and X. Tang, “A novel fault location method based on traveling wave for multi-branch distribution network,”Electr . Power Syst. Res., vol. 224, p. 109753, 2023

    2023

  10. [10]

    Ap- plication of graph neural network for fault location in PV penetrated distribution grids,

    M. MansourLakouraj, R. Hossain, H. Livani, and M. Ben-Idris, “Ap- plication of graph neural network for fault location in PV penetrated distribution grids,” inProceedings of the North American Power Sym- posium (NAPS 2021), (College Station, TX, USA), pp. 1–6, IEEE, 2021

    2021

  11. [11]

    Fault detection, classification and location for transmission lines and distribution systems: a review on the methods,

    K. Chen, C. Huang, and J. He, “Fault detection, classification and location for transmission lines and distribution systems: a review on the methods,”High V olt., vol. 1, no. 1, pp. 25–33, 2016

    2016

  12. [12]

    Spatial-temporal recurrent graph neural networks for fault diagnostics in power distribution systems,

    B. L. H. Nguyen, T. V . Vu, T.-T. Nguyen, M. Panwar, and R. Hovsapian, “Spatial-temporal recurrent graph neural networks for fault diagnostics in power distribution systems,”IEEE Access, vol. 11, pp. 46039–46050, 2023

    2023

  13. [13]

    Fault location in power distribution systems via deep graph convolutional networks,

    K. Chen, J. Hu, Y . Zhang, Z. Yu, and J. He, “Fault location in power distribution systems via deep graph convolutional networks,”IEEE J. Sel. Areas Commun., vol. 38, no. 1, pp. 119–131, 2020

    2020

  14. [14]

    The graph neural network model,

    F. Scarselli, M. Gori, A. Tsoi, M. Hagenbuchner, and G. Monfardini, “The graph neural network model,”IEEE Trans. Neural Netw., vol. 20, no. 1, pp. 61–80, 2009

    2009

  15. [15]

    A review of graph neural networks and their applications in power systems,

    W. Liao, B. Bak-Jensen, J. R. Pillai, and Y . Wang, “A review of graph neural networks and their applications in power systems,”J. Mod. Power Syst. Clean Energy, vol. 10, no. 2, pp. 345–360, 2022

    2022

  16. [16]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” inProc. Int. Conf. Learn. Represent. (ICLR 2017), (Toulon, France), pp. 1–14, OpenReview.net, 2017

    2017

  17. [17]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Li`o, and Y . Ben- gio, “Graph attention networks,” inProc. Int. Conf. Learn. Represent. (ICLR 2018), (Vancouver, BC, Canada), pp. 1–12, OpenReview.net, 2018

    2018

  18. [18]

    Inductive representation learning on large graphs,

    W. L. Hamilton, R. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” inProc. 31st Conf. Neural Inf. Process. Syst. (NIPS 2017), (Long Beach, CA, USA), pp. 3345–3355, Neural Information Processing Systems Foundation, 2017

    2017

  19. [19]

    An enhanced fault localization technique for distribution networks utilizing cost-sensitive graph neural networks,

    Z. Wang, B. Huang, B. Zhou, J. Chen, and Y . Wang, “An enhanced fault localization technique for distribution networks utilizing cost-sensitive graph neural networks,”Processes, vol. 12, no. 11, p. 2312, 2024

    2024

  20. [20]

    Fault location method of distribution network based on VGAE-GraphSAGE,

    M. Fan, J. Xia, H. Zhang, and X. Zhang, “Fault location method of distribution network based on VGAE-GraphSAGE,”Processes, vol. 12, no. 10, p. 2179, 2024

    2024

  21. [21]

    Fault diagnosis in active distribution networks with renewable energy using multi-task learning and graph attention networks,

    W. Huang, P. Chen, Y . Huang, and S. Chen, “Fault diagnosis in active distribution networks with renewable energy using multi-task learning and graph attention networks,”Electr . Power Syst. Res., vol. 263, p. 113513, 2027

    2027

  22. [22]

    A GCN-GRU-based framework for voltage sag detection and early warning in distribution networks,

    S. Pan and S. Xue, “A GCN-GRU-based framework for voltage sag detection and early warning in distribution networks,”Electr . Power Syst. Res., vol. 260, p. 113353, 2026

    2026

  23. [23]

    Decision combination in multiple classifier systems,

    T. K. Ho, J. J. Hull, and S. N. Srihari, “Decision combination in multiple classifier systems,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 16, no. 1, pp. 66–75, 1994

    1994

  24. [24]

    Robustness of spatio- temporal graph neural networks for fault location in partially observable distribution grids,

    B. Karabulut, C. Manna, and C. Develder, “Robustness of spatio- temporal graph neural networks for fault location in partially observable distribution grids,”arXiv preprint arXiv:2401.12345, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    How attentive are graph attention networks?,

    S. Brody, U. Alon, and E. Yahav, “How attentive are graph attention networks?,” inProc. Int. Conf. Learn. Represent. (ICLR 2022), (Virtual Event), pp. 1–26, OpenReview.net, 2022

    2022

  26. [26]

    OpenDSS

    The Electric Power Research Institute (EPRI), “OpenDSS.” [Online]. Available: https://www.epri.com/pages/sa/opendss, 2024

    2024

  27. [27]

    PyDSS interface

    National Renewable Energy Laboratory (NREL), “PyDSS interface.” [Online]. Available: https://www.nrel.gov/grid/pydss.html, 2024

    2024

  28. [28]

    Share of energy consumption from renewable sources in europe,

    European Environment Agency, “Share of energy consumption from renewable sources in europe,” tech. rep., EEA, 2025

    2025

  29. [29]

    Fast graph representation learning with PyTorch Geometric

    M. Fey and J. E. Lenssen, “Fast graph representation learning with PyTorch Geometric.” arXiv preprint arXiv:1903.02428, 2019

    Pith/arXiv arXiv 1903