Pith. sign in

REVIEW 5 major objections 5 minor 39 references

Enhancing Power Flow Estimation with Topology-Aware Gated Graph Neural Networks

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims a gated graph neural network trained with stability-focused settings gives more accurate AC power-flow estimates than prior GNN surrogates on grids up to 1,354 buses, with R² above 0.99.

desk verdict Useful large-scale GGNN power-flow benchmark, but the headline accuracy claims are internally contradictory and the physics-embedding claims are not implemented. read the letter →

arxiv 2507.02078 v1 pith:WTVBUZCC submitted 2025-07-02 eess.SY cs.SY

Shrenik Jadhav , Birva Sevak , Srijita Das , Wencong Su , Van-Hai Bui This is my paper
classification eess.SYcs.SY
keywords gatedgraphneuralnetworkACpowerflowestimationtopologyuncertaintyIEEEtestsystemsNewton-Raphsonsurrogateloadvariationphysics-informedtraining
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 a gated graph neural network (GGNN) is a better surrogate for AC power-flow estimation than existing GNN-based surrogates. The model is trained on 12,000 solved operating points for each of four IEEE benchmark grids (30, 118, 300, and 1354 buses), with loads varied by up to ±40% and 5% of cases containing single-line outages or transformer-tap changes. On held-out test sets, the GGNN's predicted bus voltage magnitudes and angles stay close to Newton-Raphson solutions, with the paper reporting R² above 0.99 across metrics and lower RMSE/MAE than prior GNN architectures. The significance is practical: if true, a lightweight, CPU-trained graph model could speed up contingency analysis and near-real-time grid-state estimation without iterative AC solves.

What carries the argument

The load-bearing mechanism is the gated graph neural network's GRU-based message passing. At each propagation step $t$, a node aggregates neighbor messages as $m_i^{(t)} = \sum_{j \in N(i)} W_m h_j^{(t-1)}$ and feeds them together with its previous hidden state through update and reset gates ($z_i^{(t)}$, $r_i^{(t)}$), producing a new hidden state $h_i^{(t)}$. Unrolling this single gated layer $T$ times with shared weights lets each node absorb information from up to $T$ hops away without the gradient instability of many stacked plain GNN layers. The stability-focused training configuration, especially the very small learning rate and gradient clipping, is the companion mechanism that makes the recurrent update trainable on the nonconvex AC power-flow loss.

What would settle it

Run the dataset-generation pipeline again while recording the number of discarded non-convergent or infeasible samples for each IEEE system, then evaluate the GGNN separately on the most extreme surviving cases (loads near ±40% combined with an N-1 outage). If a substantial fraction—say more than 20%—of the intended envelope was discarded, the reported R² and RMSE values would not describe the full claimed operating range; if errors on the extreme retained cases are no larger than the average, the robustness claim is confirmed.

Watch

Extended reading notes

Core claim

The central claim is that replacing fixed-depth message passing with gated recurrent message passing makes a GNN power-flow surrogate both more accurate and more scalable. The proposed GGNN treats each bus as a node whose hidden state is updated by a gated recurrent unit over a fixed number of propagation steps, so information can travel across many hops while the gates suppress vanishing-gradient instability. Trained with deliberately conservative hyperparameters—a learning rate of $5\times 10^{-5}$, 10–20% dropout, $\ell^2$ weight decay of $1\times 10^{-6}$, gradient clipping, and early stopping—it reports the best average rank across MSE, RMSE, MAE, NRMSE, and R² on IEEE 30-, 118-, 300-, and 1354-bus systems, with a 1354-bus RMSE of 0.0309 p.u. and MAE of 0.0160 p.u. The paper further claims the model respects power-flow physics because operational constraints are embedded in the node features and loss.

Load-bearing premise

The accuracy claims cover only the simulated operating points that the power-flow solver could actually solve; runs that failed to converge or produced infeasible voltage profiles were discarded without reporting how many, so the stated performance may not hold for the hardest load and outage cases.

Editorial extensions

If this is right

  • If the central claim holds, a GGNN can serve as a fast surrogate for AC power flow in real-time contingency analysis, replacing many Newton-Raphson solves with one forward pass.
  • Because the 1354-bus model trains in roughly seven hours on CPU, the authors' results imply overnight retraining is feasible as grid topology and load patterns evolve.
  • The consistent ranking across four systems and five metrics suggests the GGNN generalizes across network scale without per-system architectural redesign.
  • The low out-of-bound rates reported for the 30-bus case (1.64% for voltage magnitude, 0.03% for angle) support the claim that predictions remain within operational bounds for most test states.

Reading between the lines

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

  • The paper does not report how many simulation runs failed to converge and were discarded; counting those would show how much of the claimed ±40% load envelope the accuracy numbers actually cover.
  • The discrepancy between the 30-bus scatter-plot R² (about 0.956 for voltage magnitudes) and the Section 4.5 statement of R² above 0.99 suggests the two numbers may be computed on different subsets or aggregations; a single explicit per-system metric table would settle which is the headline result.
  • If gated recurrent depth $T$ behaves like an adaptive receptive-field size, a natural testable extension is whether the GGNN can transfer to a larger unseen grid with only fine-tuning on a few solved cases.
  • Coupling this surrogate with streaming PMU data or renewable forecasts would require continual learning, which the paper lists as future work rather than a demonstrated capability.
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

5 major / 5 minor

Summary. The manuscript proposes a gated graph neural network (GGNN) surrogate for AC power-flow estimation. The model is trained on IEEE 30-, 118-, 300-, and 1354-bus systems, with load perturbations up to ±40% and a small fraction of N-1 line-outage/tap-change topology perturbations. The authors compare the GGNN against several GNN baselines using MSE, RMSE, MAE, NRMSE, and R², and claim that the GGNN consistently achieves the lowest errors and R² above 0.99, while also emphasizing stability-focused training and physical consistency. The paper reports training convergence behavior, per-system errors, a rank-based model comparison, and a discussion of limitations and future work.

Significance. If the central claims were fully supported, the paper would offer a useful practical contribution: a single lightweight GGNN that estimates AC power-flow solutions across grid sizes from 30 to 1354 buses, including topology perturbations, with reported training times of a few hours on CPU. The multi-system, multi-metric comparison against several GNN variants is a reasonable evaluation design, and the authors explicitly acknowledge limitations regarding offline training data and the absence of generator trips and dynamic phenomena. However, the reported quantitative results are internally inconsistent, and the dataset construction omits information needed to assess the robustness claims. As it stands, the evidence does not yet support the headline conclusions.

major comments (5)
  1. [§4.5 vs §3.5 and §3.4] The claim in §4.5 that 'GGNN consistently delivers the lowest prediction errors and R² scores above 0.99 on every metric examined' is directly contradicted by §3.5, which reports R² 'about 0.956' for 30-bus voltage magnitudes, and by §3.4, which reports 30-bus RMSE of 0.0223 and MAE of 0.0131. These numbers cannot all be simultaneously correct. Moreover, R² is itself a metric; saying 'R² scores above 0.99 on every metric examined (MSE, RMSE, MAE, NRMSE)' is not meaningful as written. The authors must recompute and report a single consistent set of per-metric numbers, because the superiority claim rests on these statistics.
  2. [§3.2] The dataset description states that runs 'that failed to converge or produced infeasible voltage profiles were discarded' but does not report how many samples were discarded per system. Since the central robustness claim concerns up to 40% load variation and topology perturbations, the reported accuracy applies only to the surviving, easier-to-solve operating points. The authors should report the discard rates and analyze whether the discarded cases are systematically different from the retained ones.
  3. [§3.2] Only 5% of the 12,000 samples per system contain topology perturbations, which amounts to 600 graphs per system; with 15% of those held out for testing, only about 90 topology-perturbed test examples are available per system, and across 1,991 branches in the 1354-bus case the random N-1 outages cover a tiny fraction of possible contingencies. This is too sparse to support the paper's claim of robustness under 'significant topological uncertainty' and should be acknowledged with appropriate statistical caution.
  4. [§3.4] The reported 300-bus RMSE of 0.1775 p.u. and MAE of 0.0874 are far larger than the 1354-bus RMSE of 0.0309 and MAE of 0.0160, yet the text describes both as excellent and as evidence of graceful scaling. For per-unit voltage magnitudes, an RMSE of 0.1775 p.u. is not an excellent result, and the non-monotonic behavior across system sizes contradicts the claimed scalability pattern. The authors should discuss this anomaly and provide a plausible explanation.
  5. [Abstract, §1, §5] The abstract and conclusions claim that the model 'embeds operational constraints directly into the architecture and loss function' and that the paper explores 'physics-informed self-supervised training strategies,' but the experimental section only describes a standard supervised MSE loss (Eq. 8) and no physics-informed variant is evaluated or reported. If such a variant was studied, its results must be presented; otherwise the physical-consistency claims should be removed or substantially softened.
minor comments (5)
  1. [§3.5] Section 3.5 appears before Section 3.4 in the manuscript, which makes the reading order confusing; the sections should be renumbered or reordered.
  2. [Author affiliations] The affiliation line contains the typo 'Departmernt' in both affiliations; it should read 'Department'.
  3. [Figure 2 and Eq. (6)] Figure 2 is described as using 30×12 node embeddings, but Eq. (6) defines a seven-dimensional node feature vector; the relationship between these dimensions should be clarified.
  4. [§4.5] The phrase 'R² scores above 0.99 on every metric examined (MSE, RMSE, MAE, NRMSE)' is internally confused: R² is not a score computed for each of those error metrics. This sentence should be rewritten after the metrics are recomputed consistently.
  5. [References] Reference [3] is cited as a MISO technical report without a verifiable link or report number; a publicly accessible source should be provided.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the GGNN is evaluated as a held-out regression against external pandapower Newton–Raphson targets, and the only overlapping-author citation is background context, not load-bearing.

full rationale

The paper's central claim — that the GGNN surrogate yields lower errors and higher R² than prior GNN surrogates on IEEE 30/118/300/1354-bus systems — is a supervised regression benchmark whose targets are externally computed Newton–Raphson solutions produced by pandapower (§3.2), not quantities defined by the model's own fitted parameters. The test partition (15%) is held out from training, so the reported MSE, RMSE, MAE, and R² are genuine out-of-sample predictions. The node features include initial voltage magnitude and angle (§2.2, Eq. (6)), but in the dataset-generation pipeline these are flat-start constants (V=1.0 p.u., θ=0 rad, §3.2), so the inputs do not encode the solution being predicted. The GGNN architecture is explicitly borrowed from external prior work (Li et al. [28]); the paper even states 'we did not devise new architecture for this task.' The only reference with overlapping authorship, Mohammadi et al. [22] (Bui, Su), is cited as background context for the general statement that purely data-driven predictors can violate operational limits; it does not supply any premise of the comparative result, so it is not load-bearing. No uniqueness theorem or ansatz is imported from the authors' earlier papers. The internal inconsistency flagged by the skeptic — §3.5 reports R²≈0.956 for 30-bus magnitudes while §4.5 claims 'R² scores above 0.99 on every metric examined,' and §3.4 reports RMSE=0.0223 that is hard to reconcile with either R² — is a correctness and self-consistency defect in the reported statistics, not circularity: the metrics are still computed against external targets. Similarly, the discarding of non-converged samples (§3.2) and the 'physics-informed' framing versus the plain MSE loss in Eq. (8) are evidentiary or overclaim concerns. The derivation chain (features → GGNN → held-out NR targets) does not reduce to its inputs, so the circularity burden is minimal.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim relies on standard power flow equations as ground truth, a uniform load perturbation assumption, and the unstated assumption that discarding non-converged samples does not bias results. The GGNN hyperparameters are hand-chosen and some critical architecture parameters (hidden dimension, T) are unreported. No new entities are introduced.

free parameters (6)
  • learning_rate = 5e-5
    Chosen by hand to stabilize GGNN training on the AC power flow loss surface; no theoretical justification.
  • dropout_rate = 0.1-0.2
    Applied to hidden layers to prevent overfitting; selected via initial trials.
  • weight_decay = 1e-6
    L2 penalty chosen by hand to constrain model capacity.
  • batch_size = 16
    Chosen to balance gradient noise and convergence quality.
  • ggnn_hidden_state_dimension
    Not reported in the paper; determines model capacity and affects all numerical results.
  • message_passing_steps_T
    Number of GRU unrolling steps is not reported; directly controls the receptive field for the claimed long-range dependency handling.
assumptions (5)
  • domain assumption AC power flow equations (Eq. 1-2) accurately model steady-state grid behavior.
    Used as the ground truth generation model; standard in power systems literature.
  • domain assumption Pandapower AC power flow solver returns correct solutions for generated operating points.
    All labels are taken from pandapower; no cross-validation with an independent solver is reported.
  • domain assumption Uniform ±40% load perturbation represents a realistic operating envelope.
    The paper assumes this variation covers peak and off-peak conditions; no justification from real grid data.
  • domain assumption A GGNN can learn the AC power flow mapping from node and edge features.
    This is the central modeling hypothesis; only supported by the paper's own empirical results.
  • ad hoc to paper Discarded non-converged or infeasible samples are few and do not bias the dataset.
    The paper does not report the count of discarded cases, yet the robustness claims depend on this unstated assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Power Flow Estimation with Topology-Aware Gated Graph Neural Networks." pith.science (2026). https://pith.science/paper/WTVBUZCC

@misc{pith2026250702078,
  author       = {Pith},
  title        = {Pith review of: Enhancing Power Flow Estimation with Topology-Aware Gated Graph Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WTVBUZCC}},
  note         = {Machine review of arXiv:2507.02078}
}
read the original abstract

Accurate and scalable surrogate models for AC power flow are essential for real-time grid monitoring, contingency analysis, and decision support in increasingly dynamic and inverter-dominated power systems. However, most existing surrogates fall short of practical deployment due to their limited capacity to capture long-range nonlinear dependencies in meshed transmission networks and their weak enforcement of physical laws. These models often require extensive hyperparameter tuning, exhibit poor generalization under topology changes or large load swings, and typically do not quantify uncertainty or scale well beyond a few hundred buses. To address these challenges, this paper proposes a \textit{gated graph neural network (GGNN)} surrogate for AC power-flow estimation under topological uncertainty. The model is trained across multiple IEEE benchmark networks of varying size and complexity, each incorporating randomized line contingencies and up to 40\% load variation. To improve robustness and generalization, we explore both conventional supervised learning and physics-informed self-supervised training strategies. Comparative evaluations show that the proposed GGNN consistently outperforms prior GNN-based surrogates, achieving predictions closely aligned with Newton--Raphson solutions. By embedding operational constraints directly into the architecture and loss function, the model ensures physical consistency and delivers a lightweight, accurate, and scalable tool for real-time grid operations.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 27 canonical work pages

  1. [1]

    Power Flow Solution by Newton’s Method,

    W. F. Tinney and C. E. Hart, “Power Flow Solution by Newton’s Method,” IEEE Trans. Power App. Syst., vol. PAS‑86, no. 11, pp. 1449–1460, 1967

    work page 1967

  2. [2]

    Fast Decoupled Load Flow,

    B. Stott and O. Alsac, “Fast Decoupled Load Flow,” IEEE Trans. Power App. Syst., vol. PAS‑93, no. 3, pp. 859–869, 1974

    work page 1974

  3. [3]

    Real‑Time Contingency Analysis Performance Report,

    Midcontinent ISO, “Real‑Time Contingency Analysis Performance Report,” Tech. Rep., 2022

    work page 2022

  4. [4]

    A review on recent advances in scenario aggregation methods for power system analysis

    A. Sangadiev, A. Gonzalez-Castellanos, and D. Pozo, “A review on recent advances in scenario aggregation methods for power system analysis,” *arXiv* preprint arXiv:2207.09557, Jul. 2022

    work page Pith review arXiv 2022

  5. [5]

    Optimal Power Flow Using Graph Neural Networks,

    D. Owerko, F. Gama, and A. Ribeiro, “Optimal Power Flow Using Graph Neural Networks,” in Proc. ICASSP, 2020, pp. 5930–5934

    work page 2020

  6. [6]

    Probabilistic Power Flow Based on Physics- Guided Graph Neural Networks,

    M. Yang, G. Qiu, T. Liu, J. Liu, K. Liu, and Y. Li, “Probabilistic Power Flow Based on Physics- Guided Graph Neural Networks,” Electric Power Systems Research, vol. 235, Art. 110864, 2024, doi: 10.1016/j.epsr.2024.110864

    arXiv 2024

  7. [7]

    Graph Neural Networks for Fast Contingency Analysis of Power Systems

    A. M. Nakiganda and S. Chatzivasileiadis, “Graph Neural Networks for Fast Contingency Analysis of Power Systems,” arXiv:2310.04213, Oct. 2023

    work page Pith review arXiv 2023

  8. [8]

    PowerFlowNet: Power-flow approximation using message-passing graph neural networks,

    N. Lin, S. Orfanoudakis, N. O. Cardenas, J. S. Giraldo, and P. P. Vergara, “PowerFlowNet: Power-flow approximation using message-passing graph neural networks,” International Journal of Electrical Power & Energy Systems, vol. 160, Art. no. 110112, 2024, doi: 10.1016/j.ijepes.2024.110112

    work page arXiv 2024

Show all 39 references
  1. [9]

    Improved physics-informed neural network based AC power flow for distribution networks,

    V. Eeckhout, H. Fani, M. U. Hashmi, and G. Deconinck, “Improved physics-informed neural network based AC power flow for distribution networks,” in Proc. IEEE Power & Energy Society General Meeting (PESGM), 2024

    2024

  2. [10]

    PowerGNN: A Topology‑Aware Graph Neural Network for Electricity Grids,

    D. Suri and M. Mangal, “PowerGNN: A Topology‑Aware Graph Neural Network for Electricity Grids,” arXiv:2503.22721, Mar. 2025

    2025 arXiv

  3. [11]

    Graph Neural Networks for Efficient AC Power Flow Prediction in Power Grids,

    A. Talebi and K. Zhou, “Graph Neural Networks for Efficient AC Power Flow Prediction in Power Grids,” arXiv:2502.05702, Feb. 2025

    2025 arXiv

  4. [12]

    Physics‑Informed Graphical Neural Network for Power System State Estimation,

    Y. Zhou, J. Wang, and K. Zhao, “Physics‑Informed Graphical Neural Network for Power System State Estimation,” Appl. Energy, vol. 350, Art. 121456, 2023

    2023

  5. [13]

    Powerformer: A section-adaptive transformer for power flow adjustment,

    K. Chen, W. Luo, S. Liu, Y. Wei, Y. Zhou, Y. Qing, Q. Zhang, J. Song, and M. Song, “Powerformer: A section-adaptive transformer for power flow adjustment,” arXiv preprint arXiv:2401.02771v5, Nov. 2024

    2024 arXiv

  6. [14]

    Physics-informed graph neural network for dynamic reconfiguration of power systems,

    J. Authier, R. Haider, A. K. Annaswamy, and F. Dörfler, “Physics-informed graph neural network for dynamic reconfiguration of power systems,” arXiv preprint arXiv:2310.00728v2, Apr. 2024

    2024 arXiv

  7. [15]

    SafePowerGraph: Safety-aware evaluation of graph neural networks for transmission power grids,

    S. Ghamizi, A. Bojchevski, A. Ma, and J. Cao, “SafePowerGraph: Safety-aware evaluation of graph neural networks for transmission power grids,” arXiv preprint arXiv:2407.12421v2 [cs.LG], Jul. 2024, doi: 10.48550/arXiv.2407.12421

    2024 doi arXiv

  8. [16]

    TGACN‑TL: A Topological Graph Attention Convolution Network with Transfer Learning for Multi‑Energy Flow Calculation,

    H. Wang, X. Zhang, and Y. Liu, “TGACN‑TL: A Topological Graph Attention Convolution Network with Transfer Learning for Multi‑Energy Flow Calculation,” Energy, vol. 288, Art. 129678, 2024

    2024

  9. [17]

    A Review of Graph Neural Networks and Their Applications in Power Systems,

    Z. Liao, Y. Chen, and H. Jayasuriya, “A Review of Graph Neural Networks and Their Applications in Power Systems,” arXiv:2101.10025, 2021

    2021 arXiv

  10. [18]

    OPFData: Large-scale datasets for AC optimal power flow with topological perturbations,

    S. Lovett, M. Zgubič, S. Liguori, S. Madjiheurem, H. Tomlinson, S. Elster, C. Apps, S. Witherspoon, and L. Piloto, “OPFData: Large-scale datasets for AC optimal power flow with topological perturbations,” arXiv preprint arXiv:2406.07234, 2024

    2024 arXiv

  11. [19]

    Deep Feed‑Forward Networks for Fast Steady‑State Power Flow,

    P. Baranwal, S. Venkatasubramanian, and A. Bose, “Deep Feed‑Forward Networks for Fast Steady‑State Power Flow,” IEEE Trans. Smart Grid, vol. 10, no. 6, pp. 6206–6215, 2019

    2019

  12. [20]

    Physics-embedded graph convolutional neural network for power-flow calculation considering uncertain injections and topology,

    M. Gao, J. Yu, Z. Yang, and J. Zhao, “Physics-embedded graph convolutional neural network for power-flow calculation considering uncertain injections and topology,” IEEE Transactions on Neural Networks and Learning Systems, early access, 2023, doi: 10.1109/TNNLS.2023.3287028

    2023

  13. [21]

    Bayesian Graph Neural Networks for Power‑Flow Estimation with Uncertainty Quantification,

    J. Rivera‑Ortega, G. Lugo‑Martinez, and A. Gómez‑Expósito, “Bayesian Graph Neural Networks for Power‑Flow Estimation with Uncertainty Quantification,” Electr. Power Syst. Res., vol. 223, Art. 109459, 2023

    2023

  14. [22]

    Surrogate Modeling for Solving Optimal Power Flow: A Review,

    S. Mohammadi, V.‑H. Bui, W. Su, and B. Wang, “Surrogate Modeling for Solving Optimal Power Flow: A Review,” Sustainability, vol. 16, no. 22, Art. 9851, 2024

    2024

  15. [23]

    Power Flow Balancing with Decentralized Graph Neural Networks,

    J. B. Hansen, S. N. Anfinsen, and F. M. Bianchi, “Power Flow Balancing with Decentralized Graph Neural Networks,” IEEE Trans. Power Syst., vol. 38, no. 6, pp. 2423–2433, 2023

    2023

  16. [24]

    A. J. Wood, B. F. Wollenberg, and G. B. Sheblé, Power Generation, Operation, and Control, 3rd ed., Wiley, 2014

    2014

  17. [25]

    Ordinal optimization based security dispatching in deregulated power systems,

    Q.-S. Jia, M. Xie, and F. Wu, “Ordinal optimization based security dispatching in deregulated power systems,” in Proc. IEEE Conf. Decision and Control (CDC), 2009, pp. 6817–6822, doi: 10.1109/CDC.2009.5400740

    2009

  18. [26]

    Donon, B., Clément, R., Donnot, B., Marot, A., & Guyon, I. (2020). Neural Networks for Power Flow: Graph Neural Solver. Electric Power Systems Research, 189, 106547. https://doi.org/10.1016/j.epsr.2020.106547

    2020

  19. [27]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in Proc. Int. Conf. Learn. Representations (ICLR), 2017. [Online]. Available: https://arxiv.org/abs/1609.02907

    2017 arXiv

  20. [28]

    Gated graph sequence neural networks,

    Y. Li, D. Tarlow, M. Brockschmidt, and R. Zemel, “Gated graph sequence neural networks,” in Proc. Int. Conf. Learn. Representations (ICLR), 2016

    2016

  21. [29]

    González-Longatt, F., et al. (2022). Power Flow Analysis via Typed Graph Neural Networks. In 2022 IEEE Milan PowerTech

    2022

  22. [30]

    P., & Ba, J

    Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. Proceedings of the 3rd International Conference on Learning Representations (ICLR 2015). https://arxiv.org/abs/1412.6980

    2014 arXiv

  23. [31]

    E., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R

    Srivastava, N., Hinton, G. E., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R. R. (2014). Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15, 1929–1958

    2014

  24. [32]

    Pascanu, R., Mikolov, T., & Bengio, Y. (2013). On the difficulty of training recurrent neural networks. Proceedings of the 30th International Conference on Machine Learning (ICML 2013). https://arxiv.org/abs/1211.5063

    2013 arXiv

  25. [33]

    Wu, Z., Zhu, Y., Zhang, Z., Chen, X., & Chang, S. F. (2020). Simplifying graph convolutional networks. Proceedings of the 36th International Conference on Machine Learning (ICML 2020). https://arxiv.org/abs/2002.05287

    2020 arXiv

  26. [34]

    Paszke, A., Gross, S., & Chintala, S. (2017). Automatic differentiation in PyTorch. Proceedings of the 30th Conference on Neural Information Processing Systems (NeurIPS 2017). https://arxiv.org/abs/1611.05431

    2017 arXiv

  27. [35]

    Power Systems Test Case Archive,

    R. D. Christie, "Power Systems Test Case Archive," University of Washington, 1993. [Online]. Available: http://labs.ece.uw.edu/pstca/

    1993

  28. [36]

    AC Power Flow Data in MATPOWER and QCQP Format: iTesla, RTE Snapshots, and PEGASE,

    C. Josz and D. K. Molzahn, "AC Power Flow Data in MATPOWER and QCQP Format: iTesla, RTE Snapshots, and PEGASE," arXiv preprint arXiv:1603.01533, 2016

    2016 arXiv

  29. [37]

    Glover, J.D., Sarma, M.S., & Overbye, T.J., ‘Power System Analysis and Design,’ Cengage Learning, 2012

    2012

  30. [38]

    pandapower—An Open-Source Python Tool for Convenient Modeling, Analysis, and Optimization of Electric Power Systems,

    L. Thurner et al., “pandapower—An Open-Source Python Tool for Convenient Modeling, Analysis, and Optimization of Electric Power Systems,” IEEE Transactions on Power Systems, vol. 33, no. 6, pp. 6510–6521, 2018

    2018

  31. [39]

    James, G., Witten, D., Hastie, T., & Tibshirani, R. (2013). An Introduction to Statistical Learning: With Applications in R. Springer

    2013

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.