Pith. sign in

REVIEW 3 major objections 5 minor 30 references

Universal Graph Learning for Power System Reconfigurations: Transfer Across Topology Variations

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a single UGCN, trained on augmented reconfigurations of known grids, transfers zero-shot to unseen topologies and grid sizes without retraining, and demonstrates this on state forecasting and false-data-injection detec

desk verdict A sensible architecture for within-family reconfiguration transfer, wrapped in an 'any reconfiguration' claim the experiments don't support. read the letter →

arxiv 2509.08672 v1 pith:GYZ6EKAZ submitted 2025-09-10 eess.SY cs.SYeess.SP

Tong Wu , Anna Scaglione , Sandy Miguel , Daniel Arnold This is my paper
classification eess.SYcs.SYeess.SP
keywords UniversalGraphConvolutionalNetworkpowersystemreconfigurationzero-shottransferaugmentationadaptivepoolingparalleltransformerfalsedatainjectiondetectionstateforecasting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that one graph neural network, called the Universal Graph Convolutional Network (UGCN), can be trained on a few known power grids and then applied, with no retraining and no knowledge of the new topology, to arbitrary reconfigurations of those grids. The authors treat reconfiguration as a third learning dimension alongside space and time, generate many modified grid variants through physics-aware graph augmentation, and train shared convolution parameters across all variants. Two mechanisms make variable sizes harmless: adaptive pooling maps any number of buses to a fixed-size representation, and a parallel transformer generates outputs of whatever dimension the target configuration needs. On IEEE distribution feeders (33- and 69-bus) and transmission networks (30-, 39-, and 57-bus), the model reports large gains over baselines for both voltage-phasor forecasting and false-data-injection localization. If the claim holds, operators of reconfigured grids would not need to collect labeled data or retrain models after every topology change.

What carries the argument

The core object is the shared spatio-temporal graph filter: scalar coefficients h_{k,τ} (equivalently weight matrices H) applied to powers of the graph shift operator S_q, so the same convolution parameters work for any graph size. Around this core sit the adaptive pooling matrix A_q, which maps an arbitrary N_q-node feature matrix to N_p pooled nodes, and the parallel transformer output, which uses continuous position encodings to produce N_q outputs for any configuration. Together they remove the fixed input dimension and fixed output dimension assumptions that break ordinary neural networks on reconfigured grids.

What would settle it

Train UGCN on IEEE 33-bus augmentations, then test it on a reconfiguration generated outside the augmentation grammar—for example, a two-island microgrid operation where a feeder is intentionally islanded and resynchronized with a different phase sequence—and check whether forecasting MSE and FDI accuracy fall back to baseline levels. A second check: train on augmentations with node counts 22–38 and test on a reconfiguration with 60 nodes; if transfer degrades sharply, the 'any reconfiguration' claim fails.

Watch

Extended reading notes

Core claim

The central claim is that topology reconfiguration is not a new domain but a point along a 'topology dimension,' so a model trained on a set of augmented configurations can generalize zero-shot to any reconfiguration of the same family. The UGCN does this by sharing scalar-weight complex GCN filters across all configurations, using application-oriented adaptive pooling (average and max, or learned assignment) to collapse variable node counts to a fixed hidden size, and using a parallel transformer with position embeddings to generate outputs with configuration-specific semantics and dimensionality. The paper shows this on two tasks: state forecasting (including H=0 state estimation) and FDI

Load-bearing premise

The load-bearing premise is that the five augmentation operations—feeder disconnection, new feeder connection, parameter changes, line breaks, and subtree merging—cover every reconfiguration the model will face, so training on samples from that generator transfers to 'any' reconfiguration; the paper does not prove coverage, and its test set is drawn from the same generator.

Editorial extensions

If this is right

  • A single trained UGCN can be deployed across distribution feeder reconfigurations, such as the IEEE 33-bus system varying between 22 and 38 nodes, without retraining.
  • One model can be trained jointly on multiple heterogeneous transmission systems and still transfer to new reconfigurations of each, avoiding a separate model per grid.
  • The zero-shot transfer covers both state estimation and forecasting over horizons of 1 to 5 hours, not just single-step prediction.
  • FDI localization retains high accuracy across attack magnitudes and across multi-grid hybrid training, while baseline methods degrade toward trivial all-ones or all-zeros predictions.
  • If correct, the framework removes the need for labeled data from every new topology, making real-time redeployment after grid changes practical.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's 'any reconfiguration' claim is stronger than what the experiments directly show: the test set is drawn from the same stochastic augmentation generator as the training set, so a reconfiguration type outside that generator (for example, deliberate islanding sequences or phase-balancing operations) remains the untested frontier.
  • The same architecture could be transferred to other variable-dimension, variable-semantics power-system tasks, such as dynamic line rating under changing topology or attack localization when the sensor set changes, because the parallel transformer already decouples output semantics per configuration.
  • The scalar-shared filter view suggests that cross-system transfer between grids of very different sizes may be achievable by graph decomposition, which the paper explicitly defers; if combined with adaptive pooling, the model could learn regional physics from subgraphs and compose them for whole-grid inference.
  • A practical deployment test would compare UGCN's zero-shot predictions against a utility's actual sequence of reconfigurations and its retrained models; the paper's benchmarks do not include such a naturally occurring distribution shift.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a Universal Graph Convolutional Network (UGCN) for power system state forecasting and false data injection detection, with the aim of transferring a single trained model across reconfigurations of a given grid without retraining. The method combines physics-aware spatio-temporal graph convolutions with shared scalar filter coefficients, adaptive grid pooling to handle variable input sizes, and a parallel transformer output layer with position-based embeddings to handle variable output dimensions. Training uses graph augmentations (feeder disconnection, new feeder connection, parameter changes, line breaks, subtree merging) to generate many reconfigurations from a base grid. Experiments on IEEE 33/69-bus distribution systems and IEEE 30/39/57-bus transmission systems report large improvements over baselines in voltage forecasting and FDI localization, including multi-system hybrid training.

Significance. If the central claim of universal zero-shot transfer to any reconfiguration were established, the work would be a useful practical contribution: a single model that avoids retraining for each topology variant would address a real deployment bottleneck in power system ML. The paper's architecture is a sensible combination of existing ideas — shared spatio-temporal GCN coefficients, adaptive pooling, and a parallel transformer output — and the two application domains (state forecasting and FDI localization) are relevant and clearly formulated. The reported within-family transfer results are promising, and the paper is honest in some internal remarks (Remark 1, the conclusion) about the limits of cross-system generality. However, the 'any reconfiguration' claim in the abstract is not supported by the experimental design, and the lack of uncertainty quantification and the fragile position-encoding assumption weaken the evidence. The manuscript is likely salvageable as a within-family transfer method with substantially softened claims.

major comments (3)
  1. [Abstract; §V-A; Remark 1; §VI] The central claim of transfer to 'any reconfiguration' is not supported by the evidence. Test reconfigurations are generated by the same augmentation operations described in §III-A on the same base systems (IEEE 33/69 distribution; IEEE 30/39/57 transmission). Section V-A states 1000 training and 200 testing reconfigurations but provides no evidence that the test distribution is structurally different from the training distribution. Remark 1 explicitly defers cross-regional transfer between grids of vastly different dimensions, and the conclusion defers transfer to 'totally new grids' to future work. The experiments therefore demonstrate interpolation within an augmentation manifold, not universal extrapolation. The authors should either weaken the abstract and conclusion to 'within-family reconfiguration transfer' or add a genuinely disjoint test setting, e.g., an unseen reconfiguration
  2. [§V-D, §V-E] All reported MSE and accuracy values are point estimates with no error bars, confidence intervals, multiple seeds, or statistical significance tests. Given the very large claimed improvements (191x–271x in MSE, 27+ percentage points in accuracy), the reader cannot determine whether these gaps are robust or an artifact of a single reconfiguration draw or poorly tuned baselines. Please report mean ± standard deviation over at least five seeds and clarify whether the baselines received the same hyperparameter tuning budget and the same augmentation exposure as UGCN.
  3. [§III-D, Eq. (17)] The position encoding p_Nq = [0, ..., N_q-1]/max(N_q-1, 1) assumes that bus indices are semantically comparable across grids of different sizes. In IEEE 30/39/57-bus systems, bus numbers are conventional labels, not physical coordinates; a linear scale to [0,1] does not establish cross-grid alignment. The transformer's cross-system transfer claim depends on this assumption. Please test sensitivity to random bus permutations/renumbering to show that the reported transfer accuracy is not an artifact of bus-index alignment.
minor comments (5)
  1. [§V-D2 and Figure 7 caption] The text in §V-D2 says Figure 7 illustrates results on 'IEEE 33-bus system reconfigurations', but the figure caption reads 'unseen IEEE 69-bus system reconfiguration'. One of these is wrong; please correct.
  2. [Conclusion] There is a typo: 'beyound reconfiguration' should be 'beyond reconfiguration'. The final paragraph also repeats the same future-work sentence twice; please condense.
  3. [Remark 2] The text says 'IEEE 30, 57, and 57 bus systems' — the second '57' is likely a typo for '39'. Please fix.
  4. [Algorithm 1, line 14 vs Eq. (16)] Algorithm 1 uses mean(X_pool_q) in the encoder, while Eq. (16) uses vec(X_pool_q). These are different operations and the discrepancy should be resolved.
  5. [General] No code or data availability statement is provided. Given the complexity of the augmentation pipeline, releasing code would substantially aid reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central result is a held-out generalization test, not a fit renamed as a prediction; overclaiming in the abstract is a scope-of-claim issue, not a circularity issue.

full rationale

The paper's derivation chain is not circular by the standards of Eq. X = Eq. Y or fitted-parameter-renamed-as-prediction. The shared parameters H are trained on labeled source configurations via min_H L(H) = (1/Q) sum_q ell(H; S_q, X_q, y_q) (Eq. 8), and the reported transfer numbers (e.g., MSE 9.621e-5 in Section V-D-2, FDI accuracies in Section V-E) are evaluated on held-out labels for unseen reconfigurations; nothing in those numbers is the training objective or a fitted parameter substituted for a test target. The augmentation operations of Section III-A define the training family, and the test reconfigurations in Section V-A are drawn from the same family: '1000 distribution system reconfigurations for training and 200 for testing.' That is a standard held-out evaluation within one distribution family and is not circular. The paper's own statements limit the strongest 'any reconfiguration' claim: Remark 1 defers 'cross-regional transferability between grids with vastly different dimensions,' and the conclusion says 'Future work will extend beyond reconfigurations to achieve transfer to entirely new grids.' These are scope limitations/overclaims, which are correctness risks, not circular reductions. The self-citations [5], [21] support the GCN/GSO background but do not function as an unverified uniqueness theorem or as the sole justification for the transfer claim; the comparative experiments provide independent evidence. Therefore no circular step is present.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The ledger shows that the paper introduces no new physical entities. The main ad hoc assumption is the augmentation coverage assumption, which is the structural basis of the transfer claim. The free parameters are mostly standard hyperparameters, but the omission of N_p and the cluster assignment makes the method under-specified.

free parameters (5)
  • N_p (number of pooled nodes) = Not reported
    This hyperparameter fixes the bottleneck dimension after pooling. Its value is never stated in Section V, yet it directly controls how much topology information is retained before the output layer.
  • Graph filter order K and temporal order Kt = Not reported
    The orders of the graph and temporal convolutions in Eq. (7) are unspecified. They determine the receptive field and the amount of spatio-temporal context used.
  • Hidden dimensions and architecture widths = 256 neurons in two fully connected layers
    Standard hyperparameters chosen for the tasks, but not justified or varied.
  • Cluster assignment for adaptive average pooling = Unspecified, 'e.g., based on electrical distance or topological ordering'
    The deterministic assignment strategy is left open, so the pooling operation is not uniquely defined and results may depend on implementation choices.
  • Attack level scaling ω and detection threshold = ω in [0,1], threshold 0.5
    These application-level choices affect reported accuracy and are not systematically justified.
assumptions (4)
  • ad hoc to paper The reconfiguration types in Section III-A (feeder disconnection, new feeder, parameter changes, line breaks, subtree merging) span all reconfigurations encountered at test time, including the 'any reconfiguration' claim.
    This is the critical training-distribution coverage assumption. It is asserted, not proven, and the test reconfigurations in Section V-A are generated from the same operations.
  • domain assumption The admittance matrix Y acts as a low-pass graph filter and Y^{-1} relates voltages to currents, justifying the GSO design.
    Standard power system physics invoked in Section II-A; reasonable but not derived.
  • domain assumption Scalar (size-invariant) GCN weights transfer across graphs because graph convolutions are local with shared weights.
    This is a standard property of GCNs, adopted in Section II-B to justify cross-system transfer.
  • ad hoc to paper The position encoding p_Nq = [0,...,N_q-1]/max(N_q-1,1) provides semantically consistent bus positions across grids of different sizes.
    The assumption that linear scaling preserves bus identity semantics is not justified. It is used in Eq. (17)-(18) and is load-bearing for the parallel output layer.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Universal Graph Learning for Power System Reconfigurations: Transfer Across Topology Variations." pith.science (2026). https://pith.science/paper/GYZ6EKAZ

@misc{pith2026250908672,
  author       = {Pith},
  title        = {Pith review of: Universal Graph Learning for Power System Reconfigurations: Transfer Across Topology Variations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GYZ6EKAZ}},
  note         = {Machine review of arXiv:2509.08672}
}
read the original abstract

This work addresses a fundamental challenge in applying deep learning to power systems: developing neural network models that transfer across significant system changes, including networks with entirely different topologies and dimensionalities, without requiring training data from unseen reconfigurations. Despite extensive research, most ML-based approaches remain system-specific, limiting real-world deployment. This limitation stems from a dual barrier. First, topology changes shift feature distributions and alter input dimensions due to power flow physics. Second, reconfigurations redefine output semantics and dimensionality, requiring models to handle configuration-specific outputs while maintaining transferable feature extraction. To overcome this challenge, we introduce a Universal Graph Convolutional Network (UGCN) that achieves transferability to any reconfiguration or variation of existing power systems without any prior knowledge of new grid topologies or retraining during implementation. Our approach applies to both transmission and distribution networks and demonstrates generalization capability to completely unseen system reconfigurations, such as network restructuring and major grid expansions. Experimental results across power system applications, including false data injection detection and state forecasting, show that UGCN significantly outperforms state-of-the-art methods in cross-system zero-shot transferability of new reconfigurations.

Figures

Figures reproduced from arXiv: 2509.08672 by the authors.

Figure 1
Figure 1. Comparison of Traditional GCN and Transferable UGCN in Distribution Systems Under Different Reconfigurations. In the left figure, the trainable parameter [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. The output of the last GCN layer is processed using either average adaptive grid pooling or learnable assignment grid pooling. The former is faster and computes the average [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. UGCN framework overview: From problem formulation through universal graph [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Visualization of data manifold complexity demonstrating the universal learning challenge: (a) Traditional temporal learning within a single configuration, (b) UGCN’s ability [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: UGCN transferability for power system state estimation and forecasting across [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: UGCN transferability for power system state estimation and forecasting across [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 8
Figure 8. Figure 8: UGCN transferability for false data injection localization across different individual power system configurations without retraining. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: UGCN transferability for heterogeneous multi-grid FDI localization across [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 2 linked inside Pith

  1. [1]

    Zhu,Optimization of power system operation

    J. Zhu,Optimization of power system operation. John Wiley & Sons, 2015

    2015

  2. [2]

    Constrained reinforcement learn- ing for predictive control in real-time stochastic dynamic optimal power flow,

    T. Wu, A. Scaglione, and D. Arnold, “Constrained reinforcement learn- ing for predictive control in real-time stochastic dynamic optimal power flow,”IEEE Trans. Power Syst., vol. 39, no. 3, pp. 5077–5090, 2023

    2023

  3. [3]

    Modernizing the grid: Challenges and opportunities for a sustainable future,

    J. R. Aguero, E. Takayesu, D. Novosel, and R. Masiello, “Modernizing the grid: Challenges and opportunities for a sustainable future,”IEEE Power Energy Mag., vol. 15, no. 3, pp. 74–83, 2017

    2017

  4. [4]

    A review of safe reinforcement learning methods for modern power systems,

    T. Su, T. Wu, J. Zhao, A. Scaglione, and L. Xie, “A review of safe reinforcement learning methods for modern power systems,”arXiv preprint arXiv:2407.00304, 2024

    arXiv 2024

  5. [5]

    Complex-value spatio-temporal graph convolutional neural networks and its applications to electric power systems ai,

    T. Wu, A. Scaglione, and D. Arnold, “Complex-value spatio-temporal graph convolutional neural networks and its applications to electric power systems ai,”IEEE Trans. Smart Grid, 2023

    2023

  6. [6]

    A transfer learning framework for power system event identification,

    H. Li, Z. Ma, and Y . Weng, “A transfer learning framework for power system event identification,”IEEE Trans. Power Syst., vol. 37, no. 6, pp. 4424–4435, 2022

    2022

  7. [7]

    Transferable learning of gcn sampling graph data clusters from different power systems,

    T. Wu, A. Scaglione, D. Arnold, and T. Chen, “Transferable learning of gcn sampling graph data clusters from different power systems,” in 2024 60th Annual Allerton Conference on Communication, Control, and Computing. IEEE, 2024, pp. 1–7

    2024

  8. [8]

    Deep learning in power systems research: A review,

    M. Khodayar, G. Liu, J. Wang, and M. E. Khodayar, “Deep learning in power systems research: A review,”CSEE J. Power Energy Syst., vol. 7, no. 2, pp. 209–220, 2020

    2020

Show all 30 references
  1. [9]

    Transfer learning for transient stability predictions in modern power systems under enduring topolog- ical changes,

    M. Hijazi, P. Dehghanian, and S. Wang, “Transfer learning for transient stability predictions in modern power systems under enduring topolog- ical changes,”IEEE Trans. Autom. Sci. Eng., 2023

    2023

  2. [10]

    A decade survey of transfer learning (2010–2020),

    S. Niu, Y . Liu, J. Wang, and H. Song, “A decade survey of transfer learning (2010–2020),”IEEE Trans. Artif. Intell., vol. 1, no. 2, pp. 151– 166, 2021

    2010

  3. [11]

    V oltage stability monitoring based on disagreement-based deep learning in a time-varying environment,

    T. Wu, Y .-J. A. Zhang, and H. Wen, “V oltage stability monitoring based on disagreement-based deep learning in a time-varying environment,” IEEE Trans. Power Syst., vol. 36, no. 1, pp. 28–38, 2020

    2020

  4. [12]

    Power system event identification with transfer learning using large-scale real-world synchrophasor data in the united states,

    J. Shi, K. Yamashita, and N. Yu, “Power system event identification with transfer learning using large-scale real-world synchrophasor data in the united states,” in2022 IEEE Power & Energy Society Innovative Smart Grid Technologies Conference (ISGT). IEEE, 2022, pp. 1–5

    2022

  5. [13]

    Adaptive assessment of power system transient stability based on active transfer learning with deep belief network,

    B. Li and J. Wu, “Adaptive assessment of power system transient stability based on active transfer learning with deep belief network,” IEEE Trans. Autom. Sci. Eng., vol. 20, no. 2, pp. 1047–1058, 2022

    2022

  6. [14]

    An integrated transfer learning method for power system dynamic security assessment of unlearned faults with missing data,

    C. Ren, Y . Xu, B. Dai, and R. Zhang, “An integrated transfer learning method for power system dynamic security assessment of unlearned faults with missing data,”IEEE Trans. Power Syst., vol. 36, no. 5, pp. 4856–4859, 2021

    2021

  7. [15]

    Meta-transfer learning-based method for multi-fault analysis and assessment in power system,

    L. Zheng, Y . Zhu, and Y . Zhou, “Meta-transfer learning-based method for multi-fault analysis and assessment in power system,”Appl. Intell., vol. 54, no. 23, pp. 12 112–12 127, 2024

    2024

  8. [16]

    Meta-transfer learning for few-shot learning,

    Q. Sun, Y . Liu, T.-S. Chua, and B. Schiele, “Meta-transfer learning for few-shot learning,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 403–412

    2019

  9. [17]

    An efficient and explainable transformer-based few-shot learning for modeling electricity consumption profiles across thousands of domains,

    W. Xia, G. Peng, C. Wang, P. Palensky, E. Pauwels, and P. P. Vergara, “An efficient and explainable transformer-based few-shot learning for modeling electricity consumption profiles across thousands of domains,” arXiv preprint arXiv:2408.08399, 2024

    2024 arXiv

  10. [18]

    Transient stability assessment using deep transfer learning,

    J. Kim, H. Lee, S. Kim, S.-H. Chung, and J. H. Park, “Transient stability assessment using deep transfer learning,”IEEE Access, vol. 11, pp. 116 622–116 637, 2023

    2023

  11. [19]

    Bidirectional active transfer learning for adaptive power system stability assessment and dominant instability mode identification,

    Z. Shi, W. Yao, Y . Tang, X. Ai, J. Wen, and S. Cheng, “Bidirectional active transfer learning for adaptive power system stability assessment and dominant instability mode identification,”IEEE Trans. Power Syst., vol. 38, no. 6, pp. 5128–5142, 2022

    2022

  12. [20]

    A unified deep neural network for solving ac opf in expanding and multiple networks,

    H. Liang, C. Zhao, and M. Chen, “A unified deep neural network for solving ac opf in expanding and multiple networks,” in2024 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm). IEEE, 2024, pp. 575–580

    2024

  13. [21]

    Spatio-temporal graph convolutional neural networks for physics-aware grid learning algorithms,

    T. Wu, I. L. Carre ˜no, A. Scaglione, and D. Arnold, “Spatio-temporal graph convolutional neural networks for physics-aware grid learning algorithms,”IEEE Trans. Smart Grid, 2023

    2023

  14. [22]

    Manifold learning: What, how, and why,

    M. Meil ˘a and H. Zhang, “Manifold learning: What, how, and why,” Annual Review of Statistics and Its Application, vol. 11, no. 1, pp. 393– 417, 2024

    2024

  15. [23]

    On the challenges of learning with inference networks on sparse, high-dimensional data,

    R. Krishnan, D. Liang, and M. Hoffman, “On the challenges of learning with inference networks on sparse, high-dimensional data,” inInterna- tional conference on artificial intelligence and statistics. PMLR, 2018, pp. 143–151

    2018

  16. [24]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” inICLR, 2017

    2017

  17. [25]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Adv. Neural Inf. Process. Syst., vol. 30, 2017

    2017

  18. [26]

    Graph attention networks,

    P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Lio, Y . Bengio et al., “Graph attention networks,”stat, vol. 1050, no. 20, pp. 10–48 550, 2017

    2017

  19. [27]

    Recipe for a General, Powerful, Scalable Graph Trans- former,

    L. Ramp ´aˇsek, M. Galkin, V . P. Dwivedi, A. T. Luu, G. Wolf, and D. Beaini, “Recipe for a General, Powerful, Scalable Graph Trans- former,”Adv. Neural Inf. Process. Syst., vol. 35, 2022

    2022

  20. [28]

    Power system state forecast- ing via deep recurrent neural networks,

    L. Zhang, G. Wang, and G. B. Giannakis, “Power system state forecast- ing via deep recurrent neural networks,” inIEEE ICASSP, 2019

    2019

  21. [29]

    Locational detection of the false data injection attack in a smart grid: A multilabel classification approach,

    S. Wang, S. Bi, and Y .-J. A. Zhang, “Locational detection of the false data injection attack in a smart grid: A multilabel classification approach,”IEEE Internet Things J., vol. 7, no. 9, pp. 8218–8227, 2020

    2020

  22. [30]

    Kfrnn: an effective false data injection attack detection in smart grid based on kalman filter and recurrent neural network,

    Y . Wang, Z. Zhang, J. Ma, and Q. Jin, “Kfrnn: an effective false data injection attack detection in smart grid based on kalman filter and recurrent neural network,”IEEE Internet Things J., vol. 9, 2021

    2021

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.