Pith. sign in

REVIEW 5 major objections 5 minor 30 references

Operationally Feasible Synthetic Power-Grid Scenarios via Learning the AC-Operable Joint Distribution

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A diffusion framework that learns to sample power-grid scenarios which already pass AC power-flow checks, without post-hoc fixes.

desk verdict Plausible and clearly explained, but the feasibility gains are measured with the same solver used to train — needs an external validation before the central claim is credible. read the letter →

arxiv 2608.03878 v1 pith:23M67HUE submitted 2026-08-04 cs.LG cs.SYeess.SY

Chenhan Xiao , Xinyu He , Haoran Li , Hanghang Tong , Yang Weng This is my paper · ORCID
classification cs.LGcs.SYeess.SY
keywords syntheticpowergridsACpower-flowfeasibilityhierarchicaldiffusiongraphbetaloadprofilegenerationcontingencyrobustnessoperationallyfeasiblescenariosconstraineddistributionlearning
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

Power-system studies need large libraries of realistic grid scenarios, but statistically plausible grids often fail the nonlinear AC power-flow equations or violate voltage, thermal, and generator limits. This paper proposes to make operational feasibility an intrinsic property of the generative model rather than a post-generation filter. It factorizes the joint distribution of topology, branch parameters, and load profiles into three physically ordered stages, trains each stage with beta diffusion, and then refines the whole generator with a reward that rewards AC convergence and penalizes constraint violations. On four benchmark systems, the resulting generator samples scenarios with convergence rates around 93-100%, improved N-1 outage survival relative to baselines, and no optimization-based correction at sampling time. If correct, synthetic grids become directly usable for planning, contingency screening, and optimal-power-flow studies.

What carries the argument

The central object is the AC-operable joint distribution: a learned distribution over complete scenarios G=(X,A,E,D) that concentrates probability mass where AC power-flow equations converge and voltage, branch-flow, and generator limits hold. It is realized by a three-stage hierarchical beta-diffusion generator (topology and bus attributes, then branch parameters, then latent load trajectories), refined by a feasibility-guidance term −β E[R(G)] added to the diffusion objective. The reward R(G), computed from PYPOWER AC simulation and its constraint-violation array, is the mechanism that shifts probability mass toward operable regions while the diffusion pretraining term preserves statistica

What would settle it

Generate scenarios from a trained model and solve AC power flow under reactive-power factors drawn from a different distribution (for example, a narrower or load-dependent range), or evaluate the generator on an out-of-distribution grid not among the four benchmarks. If convergence or N-1 survival drops to the level of the random-walk baseline, or if feasibility scores change sharply with the reactive-power sampling rule, the learned AC-operability is tied to the specific reward evaluation rather than a general property of the model.

Watch

Extended reading notes

Core claim

The central claim is that AC-operable synthetic grids can be learned directly as samples of a joint distribution p(A,X,E,D) by decomposing generation along the physical hierarchy of the grid and using a simulation-based operational reward to reshape the learned distribution. The model writes the joint distribution as p(A,X) p(E|A,X) p(D|A,X,E), with beta diffusion for structure and branch parameters and an LSTM-autoencoder latent space for load trajectories. Pretraining matches the empirical data distribution; a refinement stage uses a score-function gradient estimator on the reward R(G) = λ1·1{converged} + λ2·exp(−τ‖v‖1), where v collects voltage, branch-flow, and generator-limit violations

Load-bearing premise

The central bet is that the PYPOWER-based AC convergence and constraint-violation reward, evaluated under reactive power factors sampled uniformly from [0.85,0.95], is a faithful proxy for operational feasibility beyond the training distribution.

Editorial extensions

If this is right

  • Scenario sampling needs no optimization-based post-processing: complete scenarios come directly from the diffusion reverse process, at about 0.64 seconds per sample on the 36-bus system.
  • On the four benchmarks, generated scenarios show higher AC power-flow convergence rates (98.6%, 94.5%, 92.8%, and 100%) than all compared baselines when loads are scaled and PYPOWER solves each sample.
  • N-1 outage survival improves on the transmission benchmarks (76.8% on 14-bus, 68.5% on 118-bus) and is more consistent across repeated trials, so contingency screening is less distorted by rare infeasible samples.
  • Statistical fidelity is preserved: maximum mean discrepancy scores against reference grids are low on degree, clustering, orbit, spectral, time-series, and attribute statistics, with only small trade-offs versus pure distribution-matching baselines.
  • Downstream studies, including AC optimal power flow cost distributions and load-stress/load-shedding behavior, track the reference system more closely than random-walk baselines.

Reading between the lines

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

  • The strongest test the paper does not run is out-of-distribution evaluation: applying the trained generator to an independent real grid or a benchmark not used to build the training library, preferably at a different voltage level or regulatory region, would show whether the learned AC-operability generalizes or is memorized from random-walk perturbations of the same four systems.
  • Because reactive power is assigned via power factors sampled uniformly from [0.85,0.95], the feasibility signal is tied to that sampling distribution; learning reactive-power injections jointly with the scenario, or conditioning them on generator reactive limits, is a natural extension that could further improve N-1 survival.
  • The hierarchical factorization suggests a modular transfer path: a topology module trained on one class of grids could be reused with different load-conditioning modules, enabling larger synthetic systems assembled from weakly coupled subnetworks.
  • The same feasibility-guidance reward could be attached to other generative backbones and other physics simulators, making the idea a general recipe for constrained generation of physical structures rather than a power-grid-only method.
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 paper proposes a hierarchical diffusion framework for generating synthetic power-grid scenarios, factorizing the joint distribution as p(X,A,E,D)=p_theta(A,X)p_gamma(E|A,X)p_phi(D|A,X,E). Stage S1 generates topology and bus attributes, Stage S2 generates branch parameters, and Stage S3 generates load profiles in a latent space via an LSTM autoencoder and beta diffusion. The main contribution is a feasibility-aware refinement step (Eq. 15) that adds a reward R(G)=lambda1*1_converge + lambda2*s_pf to the diffusion loss, where s_pf is an exponential penalty on PYPOWER constraint violations. The method is evaluated on IEEE 14/118/123-bus and EU 36-bus benchmarks, reporting improved AC power-flow convergence, feasibility score, N-1 survivability, and statistical fidelity compared to baselines, with no optimization-based post-processing at generation time.

Significance. If the claims hold, this is a meaningful step beyond generate-then-validate approaches: it directly biases a generative model toward AC-feasible scenarios, uses a physically motivated hierarchical decomposition whose chain-rule factorization is correct, and includes an ablation showing the feasibility guidance contributes substantially (Table IV). The claimed improvements—e.g., 98.6% convergence on IEEE 14-bus—are practically relevant. However, the current evidence is weakened by a protocol-specific evaluation: the training reward and test metric share the same PYPOWER simulation and the same random power-factor distribution, and no held-out network is tested. The manuscript also leaves the feasibility-refinement objective underspecified, which is load-bearing for the central claim.

major comments (5)
  1. [Section IV.B, Eq. (15)] The gradient update ∇_α E[R] ≈ E[R ∇_α log p_α(G)] is not implementable as stated because the paper never specifies how log p_α(G) is computed for the graph beta-diffusion model. Diffusion model log-likelihoods are generally intractable; a surrogate such as the ELBO would change the objective. Please specify the exact surrogate used, whether samples are drawn on-policy or from a replay buffer, and any variance-reduction steps. This matters because Table IV attributes a 12.1-percentage-point convergence gain to this term; without a concrete estimator, the refinement step cannot be reproduced.
  2. [Section IV.A / V.A, Eq. (13) and Table I] There is an internal inconsistency between the training signal and the evaluation metric. Section IV.A states that v is built from the 'raw constraint violation array from the runopf routine in PYPOWER', while the reported metrics are 'AC power-flow convergence' under random power factors. runopf is an OPF, not a plain power flow: it can redispatch generators, so a scenario that is infeasible under a fixed generator schedule can appear OPF-feasible. If R(G) is OPF-based and the test metric is PF-based, the two measure different physical properties; if both are OPF-based, the abstract's 'AC power-flow convergence' is overstated. The authors must clarify which solver is used for Eq. (13) and for each row of Table I, and ideally compute R with the same fixed-schedule AC power flow used at test time.
  3. [Section V.A and Table I] There is no held-out network evaluation. The training set is constructed by random-walk perturbations of the same four benchmark networks on which the model is tested (IEEE 14, EU 36, IEEE 118, IEEE 123). All reported convergence and feasibility numbers are therefore in-distribution. This does not support the claim that the model learns a general 'AC-operable joint distribution' that transfers to unseen systems. Please add leave-one-network-out experiments (e.g., train on IEEE 118/123 and test on EU 36, or train on synthetic variants of one system and test on another benchmark), or clearly restrict the claim to per-benchmark fitted generators.
  4. [Section IV.A and V.A, Eq. (13)] The training reward and the test metric assign reactive demand using random power factors uniformly sampled from [0.85,0.95] in exactly the same way. The convergence flag and feasibility score are therefore computed under the same loading protocol that the model is trained to maximize. The high scores in Table I may reflect memorization of this specific reactive-power protocol rather than robust AC operability. Please evaluate under different power-factor distributions, load scaling, solver tolerances, and at least one alternative solver, and report whether the ranking over baselines is stable.
  5. [Section V.B, baselines] The text says GDSS, GruM, and GBD receive 'the same operational feasibility guidance mechanism during training', but no details are provided for how this guidance is implemented for these models. Since Table I is the central evidence, the comparison is hard to interpret without knowing whether the same reward, the same Eq. (15) surrogate, and the same refinement schedule are used. Please specify the exact baseline training protocol or, if the baselines are not refined identically, state so explicitly and discuss the fairness implications.
minor comments (5)
  1. [Section IV.A, Eq. (12)] With tau=1e-5, s_pf is extremely insensitive to violations: even ||v||_1=1000 gives s_pf≈0.99. Please report typical violation norms or justify the chosen tau; otherwise the feasibility score may be nearly constant and the lambda2 term in Eq. (13) has little effect.
  2. [Table I] N-1 results are reported only for the two transmission systems; the abstract claims 'contingency robustness' in general. Report N-1 for EU 36 and IEEE 123 or state explicitly why the metric is not applicable to radial distribution feeders.
  3. [Table II] MMD fidelity is shown only for EU 36-bus. Include MMD for at least one additional system (e.g., IEEE 14-bus) so that the statistical-fidelity claim is not tied to a single network.
  4. [Section V.A] The random-walk perturbation procedure is described only in one sentence. For reproducibility, give the number of edge swaps, the distribution of impedance perturbations, the number of load-profile samples per topology variant, and the train/validation split.
  5. [Section III.B] Beta diffusion operates on continuous values in [0,1], but the adjacency matrix is binary. Clarify how the generated continuous edge scores are binarized (thresholding, Bernoulli sampling, or Gumbel trick) and how that affects the reported convergence metrics.

Circularity Check

2 steps flagged · score 6.0 of 10

Feasibility gains are partially fitted values: Table I's convergence rate and feasibility score are the same PYPOWER-based reward components that Eq. (15) maximizes; N–1, ACOPF, and load-stress results provide independent but partial support.

  1. fitted input called prediction [Section IV.A (Eq. 13), Section IV.B (Eq. 15), Section V.A/V.C (metrics)]
    "R( ˜G) =λ11converge( ˜G) +λ2spf ( ˜G) ... We report the convergence rate, defined as the fraction of generated samples whose AC power flow converges under the synthesized load profiles. ... Beyond binary convergence, we evaluate operational validity using the feasibility score (Section IV-A)."

    The training objective in Eq. (15) maximizes E_{G∼pα}[R(G)] using the score-function gradient, so the model parameters are explicitly fitted to increase λ1·1converge + λ2·spf. The evaluation then reports the same two quantities: convergence rate is the empirical mean of 1converge and the feasibility score is spf, both computed with the same PYPOWER simulation and the same random power-factor sampling (uniform [0.85,0.95]) used to compute the reward. Table I's Conv. and Feas. columns are therefore not independent predictions of AC-operability; they are the training objective evaluated on new samples. The N–1 and ACOPF analyses are not part of R(G) and provide some independent evidence, which is why the circularity is partial rather than total.

  2. other [Section IV.A (PYPOWER runopf usage); Section V.A (feasibility-score metric)]
    "In experiments, v is constructed using the raw constraint violation array from the runopf routine in PYPOWER [27], accessed via results["raw"]["g"]."

    This sentence follows the claim that each scenario is 'evaluated through AC power-flow simulation,' but runopf is an optimal-power-flow routine, not a plain AC power-flow solver. If the reward and the reported feasibility score are computed from the OPF solution, a generated scenario can be scored as feasible because the OPF redispatched generation or adjusted controls, not because the sampled topology/parameter/load combination itself admits an AC power-flow solution. The evaluation metric is then produced by the same solver that supplies the training signal, so the high feasibility scores partly certify the OPF routine's repair capability rather than the generator's raw AC-operability. This compounds the reward/evaluation coupling and undercuts the abstract's claim of eliminating optimiz

full rationale

The paper's central derivation (hierarchical factorization + beta diffusion + latent load model) is not circular: the factorization is a standard conditional decomposition, the beta-diffusion backbone is cited to non-overlapping external work, and no uniqueness theorem or load-bearing self-citation is invoked (PowerGrow appears only as a baseline and as prior work). The circularity is concentrated in the operational-feasibility claim. Eq. (15) explicitly trains the model to maximize R(G)=λ1·1converge+λ2·spf, and the evaluation metrics in Table I are exactly the empirical means of these two terms, computed under the same PYPOWER protocol and the same uniform [0.85,0.95] power-factor sampling. Thus the reported convergence rate and feasibility score are fitted values, not out-of-sample predictions. The runopf inconsistency makes this worse: the feasibility score and convergence signal are taken from an OPF routine rather than a plain power-flow solve, so the metric may credit scenarios that are only OPF-repairable. Independent content does exist: N–1 outage survivability, ACOPF cost distributions, load-stress shedding, MMD fidelity, and sampling-time comparisons are not part of the training reward and are therefore non-circular evidence; the N–1 and ACOPF results in particular support the claim of improved robustness and economic realism. However, the headline 'significantly improves operational feasibility' rests mainly on the reward-coupled metrics, and the absence of a held-out network evaluation means generalization is untested. These are correctness/generalization risks and are not counted as additional circularity. Overall, the operational-feasibility result partially reduces by construction to its training objective, warranting a 6.

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

The central claim rests on the validity of the PYPOWER-based reward signal, the representativeness of the random-walk training data, and the tractability of the REINFORCE-style gradient on a diffusion model. None of these are supplied with independent evidence. No new physical entities are introduced.

free parameters (5)
  • tau (feasibility score temperature) = 1e-5
    Controls the exponential mapping from violation norm to score in Eq. (12); chosen by hand without sensitivity analysis.
  • lambda1, lambda2 (reward weights) = 0.5 each
    Weights balancing convergence and feasibility terms in Eq. (13); no ablation on these weights reported.
  • beta (guidance strength) = 0.01
    Scales the reward term in the total loss Eq. (15); chosen by hand.
  • latent dimension d4 = 64
    Dimension of the LSTM autoencoder latent space for load profiles; no analysis of its effect on feasibility.
  • power factor range = [0.85, 0.95]
    Uniform sampling range for reactive power assignment in AC simulation; the feasibility guarantee is conditional on this range.
assumptions (4)
  • domain assumption PYPOWER AC power-flow solver provides ground-truth feasibility assessment
    The training reward and all operational metrics are computed from PYPOWER; no independent verification against real grid operation.
  • domain assumption Random-walk perturbations of benchmark systems yield a representative training distribution for real grid scenarios
    The scenario library is built from degree-preserving edge swaps and parameter perturbations of the same benchmarks later used for evaluation (Section V.A).
  • domain assumption The score-function gradient estimator is unbiased and computable with log p_alpha of the beta-diffusion model
    Eq. (15) uses E[R grad log p_alpha]; the paper does not detail how log p_alpha is obtained for the combined discrete-continuous diffusion model.
  • domain assumption LSTM autoencoder latent space of dimension 64 preserves feasibility-relevant load dynamics
    Section III.C compresses T-step trajectories to 64 dims without evidence that feasibility-relevant extremes are retained.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Operationally Feasible Synthetic Power-Grid Scenarios via Learning the AC-Operable Joint Distribution." pith.science (2026). https://pith.science/paper/23M67HUE

@misc{pith2026260803878,
  author       = {Pith},
  title        = {Pith review of: Operationally Feasible Synthetic Power-Grid Scenarios via Learning the AC-Operable Joint Distribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/23M67HUE}},
  note         = {Machine review of arXiv:2608.03878}
}
read the original abstract

Synthetic power-grid scenarios are essential for planning, resilience assessment, contingency analysis, and data-driven power-system applications. Recent synthetic grid generation methods have improved structural realism and operational feasibility by incorporating engineering knowledge through post-generation validation, optimization, or physics-aware generation. However, generated scenarios may still exhibit low AC feasibility and robustness, limiting their practical value for downstream power-system studies. This paper proposes a feasibility-aware distribution-learning framework that learns the AC-operable joint distribution of network topology, branch electrical parameters, and time-varying load profiles. Instead of enforcing feasibility after generation, the proposed framework incorporates AC power-flow convergence and operational constraints into hierarchical diffusion-based distribution learning. This enables the generator itself to produce operationally feasible grid scenarios through efficient diffusion sampling. The hierarchical architecture decomposes the high-dimensional generation task into three engineering-motivated stages: topology and bus-attribute generation, branch-parameter generation conditioned on the generated structure, and load-profile generation conditioned on both network structure and electrical characteristics. Experiments on benchmark systems demonstrate that the proposed framework significantly improves operational feasibility and contingency robustness while maintaining strong statistical fidelity and eliminating optimization-based post-processing.

Figures

Figures reproduced from arXiv: 2608.03878 by the authors.

Figure 1
Figure 1. Hierarchical factorization of the structural-operational distribution. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Violin plots of distribution of operational metrics for generated IEEE [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Original IEEE 14-bus transmission network (left) and example [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Original EU 36-bus medium-voltage urban distribution network (left) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Top: normalized historical bus-level active power trajectories used for [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: Evolution of generated topology across training epochs in the EU [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Evolution of generated bus-level load trajectories during training in [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Histogram of ACOPF costs for grids generated by the proposed model [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Load-shedding fraction against load scaling factor [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 28 canonical work pages

  1. [1]

    Powergrow: feasible co-growth of structures and dynamics for power grid synthesis,

    X. He, C. Xiao, H. Li, R. Qiu, Z. Xu, Y . Weng, J. He, and H. Tong, “Powergrow: feasible co-growth of structures and dynamics for power grid synthesis,” inProceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V . 1, 2026, pp. 428–439

    work page 2026

  2. [2]

    Distribution grid topology and parameter estimation using deep-shallow neural network with physical consistency,

    H. Li, Y . Weng, V . Vittal, and E. Blasch, “Distribution grid topology and parameter estimation using deep-shallow neural network with physical consistency,”IEEE Transactions on Smart Grid, vol. 15, no. 1, pp. 655– 666, 2023

    work page 2023

  3. [3]

    Real-time Assessments,

    C. Calhoun and B. Thiry, “Real-time Assessments,” Federal Energy Regulatory Commission and ERO Enterprise, Tech. Rep., Aug. 2021. [Online]. Available: https://www.rfirst.org/wp-content/uploads/2023/09/ Real-time-Assessments-Joint-FERC-ERO-Paper-August-2021.pdf

    work page 2021

  4. [4]

    Transmission Operator Workflows for Real-Time Reliability Studies,

    A. A. Anderson, S. Kincic, C. K. Fallon, and B. A. Jefferson, “Transmission Operator Workflows for Real-Time Reliability Studies,” Pacific Northwest National Laboratory, Tech. Rep. PNNL-36516, Oct. 2024. [Online]. Available: https://www.pnnl.gov/main/publications/ external/technical reports/PNNL-36516.pdf

    work page 2024

  5. [5]

    Distribution grid line outage iden- tification with unknown pattern and performance guarantee,

    C. Xiao, Y . Liao, and Y . Weng, “Distribution grid line outage iden- tification with unknown pattern and performance guarantee,”IEEE Transactions on Power Systems, 2023

    work page 2023

  6. [6]

    Real-Time Contingency Analysis With Corrective Transmission Switching,

    X. Li, P. Balasubramanian, M. Sahraei-Ardakani, M. Abdi-Khorsand, K. W. Hedman, and R. Podmore, “Real-Time Contingency Analysis With Corrective Transmission Switching,”IEEE Transactions on Power Systems, vol. 32, no. 4, pp. 2604–2617, 2017

    work page 2017

  7. [7]

    Fault Location, Isolation, and Service Restoration Technologies Reduce Outage Impact and Duration,

    U.S. Department of Energy, “Fault Location, Isolation, and Service Restoration Technologies Reduce Outage Impact and Duration,” Tech. Rep., Dec. 2014. [Online]. Available: https://www.energy.gov/sites/ prod/files/2016/10/f33/Fault Location Impact Duration Dec 2014.pdf

    work page 2014

  8. [8]

    Low-dimensional ode embedding to convert low-resolution meters into “virtual

    H. Li, Z. Ma, Y . Weng, H. Zhong, and X. Zheng, “Low-dimensional ode embedding to convert low-resolution meters into “virtual” pmus,” IEEE Transactions on Power Systems, vol. 40, no. 2, pp. 1439–1451, 2024

    work page 2024

Show all 30 references
  1. [9]

    A frame- work for generating synthetic distribution feeders using openstreetmap,

    S. S. Saha, E. Schweitzer, A. Scaglione, and N. G. Johnson, “A frame- work for generating synthetic distribution feeders using openstreetmap,” in2019 North American Power Symposium (NAPS). IEEE, 2019, pp. 1–6

    2019

  2. [10]

    Synthetic grid generator: Synthesizing large- scale power distribution grids using open street map,

    C. S. C. Dande, L. Mattorolo, J. d. S. Andre, L. Lavecchia, N. Efkar- pidis, and D. Toffanin, “Synthetic grid generator: Synthesizing large- scale power distribution grids using open street map,”arXiv preprint arXiv:2408.04923, 2024

    2024 arXiv

  3. [11]

    Generation of low-voltage synthetic grid data for energy system modeling with the pylovo tool,

    B. R. Baecker, S. Candas, D. Tepe, and A. Mohapatra, “Generation of low-voltage synthetic grid data for energy system modeling with the pylovo tool,”Sustainable Energy, Grids and Networks, vol. 41, p. 101617, 2025

    2025

  4. [12]

    Grid structural characteristics as validation criteria for synthetic networks,

    A. B. Birchfield, T. Xu, K. M. Gegner, K. S. Shetye, and T. J. Over- bye, “Grid structural characteristics as validation criteria for synthetic networks,”IEEE Transactions on power systems, vol. 32, no. 4, pp. 3258–3265, 2016

    2016

  5. [13]

    Statistical considerations in the creation of realistic synthetic power grids for geomagnetic disturbance studies,

    A. B. Birchfield, K. M. Gegner, T. Xu, K. S. Shetye, and T. J. Overbye, “Statistical considerations in the creation of realistic synthetic power grids for geomagnetic disturbance studies,”IEEE Transactions on Power Systems, vol. 32, no. 2, pp. 1502–1510, 2016

    2016

  6. [14]

    Generating synthetic power grids using exponential random graph models,

    F. Giacomarra, G. Bet, and A. Zocca, “Generating synthetic power grids using exponential random graph models,”PRX Energy, vol. 3, no. 2, p. 023005, 2024

    2024

  7. [15]

    A methodology for the creation of geographically realistic synthetic power flow models,

    K. M. Gegner, A. B. Birchfield, T. Xu, K. S. Shetye, and T. J. Overbye, “A methodology for the creation of geographically realistic synthetic power flow models,” in2016 IEEE Power and Energy Conference at Illinois (PECI). IEEE, 2016, pp. 1–6

    2016

  8. [16]

    A random growth model for power grids and other spatially embedded infrastructure networks,

    P. Schultz, J. Heitzig, and J. Kurths, “A random growth model for power grids and other spatially embedded infrastructure networks,”The European Physical Journal Special Topics, vol. 223, no. 12, pp. 2593– 2610, 2014

    2014

  9. [17]

    A two-stage ai- powered motif mining method for efficient power system topological analysis,

    Y . Li, Z. Zhou, J. Ping, X. Xu, Z. Yan, and J. Wu, “A two-stage ai- powered motif mining method for efficient power system topological analysis,”Applied Energy, vol. 416, p. 127992, 2026

    2026

  10. [18]

    Synthetic active dis- tribution system generation via unbalanced graph generative adversarial network,

    R. Yan, Y . Yuan, Z. Wang, G. Geng, and Q. Jiang, “Synthetic active dis- tribution system generation via unbalanced graph generative adversarial network,”arXiv preprint arXiv:2108.00599, 2021

    2021 arXiv

  11. [19]

    Deep generative graph distribution learning for synthetic power grids,

    M. Khodayar, J. Wang, and Z. Wang, “Deep generative graph distribution learning for synthetic power grids,”arXiv preprint arXiv:1901.09674, 2019

    1901 arXiv

  12. [20]

    Feedergan: Synthetic feeder generation via deep graph adversarial nets,

    M. Liang, Y . Meng, J. Wang, D. L. Lubkeman, and N. Lu, “Feedergan: Synthetic feeder generation via deep graph adversarial nets,”IEEE Transactions on Smart Grid, vol. 12, no. 2, pp. 1163–1173, 2020

    2020

  13. [21]

    Exploring variational graph au- toencoders for distribution grid data generation,

    S. Z. Abbas and E. Okoyomon, “Exploring variational graph au- toencoders for distribution grid data generation,”arXiv preprint arXiv:2509.02469, 2025

    2025

  14. [22]

    Score-based generative modeling of graphs via the system of stochastic differential equations,

    J. Jo, S. Lee, and S. J. Hwang, “Score-based generative modeling of graphs via the system of stochastic differential equations,” inInterna- tional conference on machine learning. PMLR, 2022, pp. 10 362– 10 383

    2022

  15. [23]

    Advancing graph generation through beta diffusion,

    X. Liu, Y . He, B. Chen, and M. Zhou, “Advancing graph generation through beta diffusion,” inThe Thirteenth International Conference on Learning Representations, 2025. [Online]. Available: https://openreview.net/forum?id=x1An5a3U9I

    2025

  16. [24]

    Active distribution system synthesis via unbalanced graph generative adversarial network,

    R. Yan, Y . Yuan, Z. Wang, G. Geng, and Q. Jiang, “Active distribution system synthesis via unbalanced graph generative adversarial network,” IEEE Transactions on Power Systems, vol. 38, no. 5, pp. 4293–4307, 2022

    2022

  17. [25]

    Beta diffusion,

    M. Zhou, T. Chen, Z. Wang, and H. Zheng, “Beta diffusion,”Advances in Neural Information Processing Systems, vol. 36, pp. 30 070–30 095, 2023

    2023

  18. [26]

    Digress: Discrete denoising diffusion for graph generation,

    C. Vignac, I. Krawczuk, A. Siraudin, B. Wang, V . Cevher, and P. Frossard, “Digress: Discrete denoising diffusion for graph generation,” inICLR, 2023

    2023

  19. [27]

    PYPOWER: Electric power system simulation in python,

    R. Richards, “PYPOWER: Electric power system simulation in python,” https://github.com/rwl/PYPOWER, 2013, version 5.0

    2013

  20. [28]

    European representative electricity distribution networks,

    C. Mateo, G. Prettico, T. G ´omez, R. Cossent, F. Gangale, P. Fr ´ıas, and G. Fulli, “European representative electricity distribution networks,” International Journal of Electrical Power & Energy Systems, vol. 99, pp. 273–280, 2018

    2018

  21. [29]

    A gnn-based generative model for generating synthetic cyber-physical power system topology,

    Y . Liu, H. Xie, A. Presekal, A. Stefanov, and P. Palensky, “A gnn-based generative model for generating synthetic cyber-physical power system topology,”IEEE Transactions on Smart Grid, vol. 14, no. 6, pp. 4968– 4971, 2023

    2023

  22. [30]

    Graph generation with diffusion mixture,

    J. Jo, D. Kim, and S. J. Hwang, “Graph generation with diffusion mixture,” inInternational Conference on Machine Learning. PMLR, 2024, pp. 22 371–22 405

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.