Pith. sign in

REVIEW 3 major objections 5 minor 47 references

Constrained Diffusion Models for Synthesizing Representative Power Flow Datasets

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Adding a physics-based guidance term to diffusion sampling yields synthetic power-flow datasets that are both more feasible and closer to the ground-truth distribution than unguided diffusion.

desk verdict A useful empirical contribution to constrained diffusion for power flow data, with a proof that does not stand up to scrutiny as written. read the letter →

arxiv 2506.11281 v2 pith:M4M3S5CR submitted 2025-06-12 cs.LG cs.SYeess.SY

Milad Hoseinpour , Vladimir Dvorkin This is my paper
classification cs.LGcs.SYeess.SY
keywords diffusionmodelspowerflowsyntheticdatagenerationgradientguidanceconstraintfeasibilitymanifoldlearningACoptimalphysics-informedmachine
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 argues that a diffusion model can generate synthetic power-flow datasets that are both statistically representative of real operating records and feasible under AC power-flow constraints, provided sampling is steered by a gradient-guidance term built from the constraint residual. The central claim is that this guidance moves each denoised estimate along the clean data manifold toward the constraint-satisfying set instead of distorting the learned distribution. If true, system operators could release physics-consistent synthetic grid data for downstream machine-learning tasks without exposing raw operational records. Experiments on three IEEE benchmark systems show that guided sampling reduces active and reactive power mismatches and lowers the Wasserstein distance to the ground-truth dataset.

What carries the argument

The central object is the manifold-constrained gradient guidance term: at each reverse step the sampler estimates the clean sample by denoising, then corrects it by descending the squared AC power-flow residual along $\nabla_{x_t} R_H(\hat{x}_{0|t})$, scaled by a step size $\lambda_t$. The enabling identity is Theorem 1: under the local-affine approximation of the clean data manifold, this gradient is tangential to the manifold, so no explicit tangent-space projection is needed. Two implementation devices carry the practical gains: variable decoupling, which splits the state into $(p,\theta)$ and $(q,v)$ and trains two smaller denoisers after the pattern of fast-decoupled power flow, and min-max normalization with chain-rule gradient rescaling so the guidance step acts consistently across variables of different physical scale.

What would settle it

Compute, on a benchmark grid, both the raw gradient correction $\nabla_{x_t} R_H(\hat{x}_{0|t})$ and the explicitly Riemannian projected gradient $\mathrm{Proj}_{T_{\hat{x}_{0|t}}\mathcal{M}}\nabla_{x_t} R_H(\hat{x}_{0|t})$; if repeated sampling shows they lead to materially different distributions, Theorem 1's tangential-gradient conclusion is violated. A complementary test: increase $\lambda_t$ until power-balance mismatches vanish and check whether the Wasserstein distance to ground truth grows relative to unguided sampling, which would falsify the claim that feasibility gains come without statistical loss.

Watch

Extended reading notes

Core claim

The paper's core discovery is that a one-step correction $\hat{x}'_{0|t} = \hat{x}_{0|t} - \lambda_t \nabla_{x_t} R_H(\hat{x}_{0|t})$, applied to the diffusion denoiser's clean estimate with $R_H(x)=\|H(x)\|_2^2$ the AC power-flow residual, drives synthetic samples toward the intersection of the constraint set and the clean data manifold. Theorem 1 asserts that under a local affine-subspace assumption the gradient of the residual already lies in the tangent space of the manifold, so the projection step of Riemannian gradient descent is unnecessary. In the reported experiments, guided sampling yields near-zero power-balance mismatches at most buses and also reduces the Wasserstein distance between synthetic and ground-truth datasets, so feasibility and statistical representativeness improve together. Models trained on guided synthetic data also give smaller power-flow mismatches in a downstream warm-start prediction task than models trained on unguided synthetic data.

Load-bearing premise

The no-projection argument rests on assuming that the denoiser's clean estimate lies exactly on the manifold of physically valid power-flow states and behaves locally as an orthogonal projection onto that manifold; the statistical-similarity results also assume that AC-OPF solutions generated from uniformly perturbed loads stand in for real operational data.

Editorial extensions

If this is right

  • Synthetic datasets inherit the statistical properties of the training records, including multi-modal joint distributions, while satisfying AC power-flow balance constraints at most buses.
  • Gradient guidance lowers the type-1 Wasserstein distance between synthetic and ground-truth datasets relative to unguided diffusion across the PJM 5-bus, IEEE 24-bus, and IEEE 118-bus systems.
  • Neural networks trained on guided synthetic data produce smaller active and reactive power mismatches in a downstream power-flow warm-start task than those trained on unguided data.
  • Variable decoupling between $(p,\theta)$ and $(q,v)$ allows training and sampling with two smaller denoisers, improving scalability without sacrificing the reported gains.
  • System operators can use the framework to publish physics-consistent synthetic power-flow data for ML applications without revealing raw operational records.

Reading between the lines

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

  • The representativeness claims are measured against a ground truth built by AC-OPF on uniformly perturbed loads, so the natural next test is whether the same statistical-similarity and feasibility gains survive on genuine historical utility records.
  • The guidance recipe should transfer to any domain where a smooth residual function defines physical feasibility and a diffusion model learns the feasible states, such as gas or water network flow, as long as the tangential-gradient property holds.
  • A direct stress test would be to replace the single gradient step with several projected-gradient iterations per reverse step, checking whether residuals can be pushed toward machine precision without degrading statistical similarity.
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 diffusion-model framework for generating synthetic AC power flow datasets that are both statistically similar to a given historical dataset and feasible with respect to the AC power flow equations. The key methodological contribution is a gradient guidance term added during reverse diffusion sampling, based on the residual of the power flow equations, which is claimed to steer samples toward the intersection of the feasible set and the clean data manifold without projection. The authors also introduce a variable decoupling strategy inspired by fast decoupled power flow and a min-max normalization scheme. Experiments on IEEE 5-, 24-, and 118-bus systems show that guided sampling reduces power mismatch and yields slightly lower Wasserstein distances than unguided sampling, and that models trained on guided samples produce better power-flow warm starts.

Significance. If the method works as claimed, it addresses a real need for privacy-preserving, physics-consistent synthetic grid data for downstream ML applications. The paper is commendable for shipping a concrete implementation with algorithms, a public code repository, and experiments across three benchmark systems, including a downstream utility test. The empirical comparison (guided vs. unguided) is consistent, and the improvement in feasibility is clearly demonstrated. However, the theoretical guarantee that guidance does not push samples off the learned distribution is not proven as stated, and several claims about statistical similarity are confounded by the feasibility effect. The central practical contribution is defensible, but the formal overclaim needs to be repaired or carefully repositioned.

major comments (3)
  1. [Sec. IV, Theorem 1, Eqs. (18)-(23)] The proof of Theorem 1 is not valid for a trained denoising network. The assumption imported from Proposition 2 of [40] is that Q is an orthogonal projection onto the clean data manifold, i.e., Q(x_t) ∈ M and JQ = JQ^T = JQ^2. For a trained diffusion model, Q(x_t) is the Tweedie posterior mean E[x_0 | x_t], which is not idempotent and does not satisfy the projection property at finite noise levels. Moreover, Eq. (18) computes the gradient with respect to the noisy input x_t, while the update (23) modifies the clean estimate \hat{x}_0; these two gradients can be identified only through JQ, and that identification is exactly what the projection assumption supplies. Consequently, the theorem does not establish that the guidance direction is tangential to M, and the claimed guarantee that the final sample 'remains on the clean data manifold' is unsupported. The authors should either prove the result under weaker assumptions or clearly label the no-projection guidance as a heuristic approximation and remove the phrase 'formally prove' from Contribution 2.
  2. [Sec. IV, after Eq. (23), and Fig. 3b] The statement that a tangential guidance step keeps \hat{x}'_{0|t} on the clean data manifold is incorrect for a curved manifold. Even if ∇ R_H(\hat{x}_{0|t}) lies in the tangent space T_{\hat{x}_{0|t}} M, the update \hat{x}' = \hat{x} - λ_t ∇ R_H is taken in the ambient Euclidean space; for a nonlinear manifold this update leaves the manifold at second order in λ_t. The local affine subspace assumption in Theorem 1 cannot justify the global conclusion drawn in the text. The paper should restrict the claim to a first-order differential statement or add a retraction/projection step that explicitly keeps the updated sample on M.
  3. [Sec. VI-B, Table I and subsequent discussion] The conclusion that 'constraint enforcement enhances the statistical similarity' is confounded by the fact that the ground truth lies exactly on the feasible set. Any reduction in constraint violations will mechanically reduce the Wasserstein distance to the ground truth, because the guided samples are moved closer to the feasible set where all ground-truth points reside. The small W1 improvements are therefore expected and do not by themselves demonstrate that guidance improves distributional fidelity beyond the feasibility effect. Please provide an evaluation that isolates distributional similarity, such as conditional distributions or standardized moments of selected variables, to support the claim.
minor comments (5)
  1. [References, [21]] Reference [21] cites 'Domain-constrained diffusion models to synthesize tabular data: A case study in power systems' with the same arXiv number as the present manuscript (2506.11281); if this is a different paper, the arXiv identifier is likely incorrect, and if it is the same paper, the title is inconsistent. Please correct the reference.
  2. [Sec. V-A] The variable decoupling trains two independent denoisers for (p, θ) and (q, v), implicitly assuming conditional independence between these groups; this may reduce the ability to capture cross-group correlations. Please discuss the impact or provide a quantitative comparison against a jointly trained model on at least one test case.
  3. [Algorithms 3 and 5] The output variable \tilde{x}_0 is never explicitly assigned in either algorithm; the algorithms should state that the return value is the denoised sample at t = 0.
  4. [Sec. VI-A] The ground-truth dataset is generated from uniform load perturbations and AC-OPF solutions; the abstract and conclusion claim that the model synthesizes data from 'real-world power grids,' but the experiments use benchmark systems with a synthetic ground truth. Please state this limitation more explicitly.
  5. [Sec. IV, Eq. (23)] The guidance scale λ_t is a free hyperparameter, and no tuning procedure is given beyond the values reported in the figures and tables. A brief sensitivity analysis or a heuristic for choosing λ_t would improve reproducibility.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity: feasibility gains are the designed objective and the comparison is against an external unconstrained baseline; the only self-referential item is a minor citation to the paper's own arXiv ID.

  1. other [Section I (Related Work) and References, item [21]]
    "“statistical consistency alone cannot guarantee adherence to physical grid constraints [21].” Reference [21] reads: “M. Hoseinpour and V . Dvorkin, ‘Domain-constrained diffusion models to synthesize tabular data: A case study in power systems,’ arXiv preprint arXiv:2506.11281, 2025.”"

    The citation’s arXiv identifier is identical to this manuscript’s identifier (2506.11281), so the paper supports a motivating premise by citing itself under a different title. The premise—that statistical matching alone does not enforce physics—is independently demonstrated by the paper’s own unconstrained-diffusion baseline, and the method’s evaluation does not depend on this citation. It is therefore a minor self-referential citation rather than a load-bearing circular step.

full rationale

The central claim is the design of a physical-constraint guidance term in diffusion sampling, benchmarked against unconstrained diffusion. The guidance term explicitly performs gradient descent on the AC power flow residual ∥H∥², so the observed feasibility improvement is the intended effect of the algorithm rather than an independent prediction; that is by-design optimization, not circularity. Statistical-similarity improvement is evaluated by Wasserstein distance against ground truth generated by AC-OPF; because the ground truth lies on the constraint set, moving samples toward the constraint set reasonably reduces W1, but the paper does not fit any parameter to the W1 metric and the comparison to the unconstrained baseline remains a valid external benchmark. Theorem 1's no-projection conclusion rests on Proposition 2 of Chung et al. [40], namely that the denoising map Q behaves locally as an orthogonal projection onto the data manifold. That is an imported external assumption, not a self-citation, and it is not an equation-level reduction of the paper's own definitions; whether trained denoisers actually satisfy it is a correctness/validity concern, not circularity. The only genuinely self-referential item is reference [21], which cites the current manuscript's own arXiv ID under a different title to support a motivating premise; this has no effect on the numerical comparisons and is therefore scored as one minor, non-load-bearing self-citation.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical entities. Its main load-bearing imports are the manifold projection heuristic from Chung et al. and the synthetic ground truth from AC-OPF, plus the tuned guidance scale lambda.

free parameters (1)
  • guidance scale lambda_t = 1e-2 (5-bus), 1e-4 (24-bus), 5e-4 (118-bus)
    A hyperparameter that controls the strength of the constraint-guidance step. It is tuned per test case to balance feasibility against distribution distortion, with no stated selection rule or sensitivity analysis.
assumptions (5)
  • domain assumption The denoiser's clean estimate Q(x_t) locally behaves as an orthogonal projection onto the clean data manifold, with JQ = JQ^T = JQ^2 (Proposition 2 of [40]).
    Used in Theorem 1, eqs. (19)-(20), to conclude the residual gradient is tangent and projection is unnecessary. This is a heuristic property of diffusion posterior sampling, not a proven fact for arbitrary trained denoisers.
  • domain assumption The clean data manifold is locally well approximated by an affine subspace around each estimate.
    Stated before Theorem 1 as the condition for eq. (15); it is the premise that lets the paper drop the Riemannian projection operator.
  • domain assumption Ground truth datasets generated by uniform load perturbations around nominal values and AC-OPF solutions are representative of real-world power flow records.
    Section VI-A states real-world data are unavailable; all statistical-similarity numbers are relative to this synthetic ground truth, so transfer to real operations is assumed.
  • domain assumption Active power is strongly correlated with voltage angles and reactive power with voltage magnitudes, justifying the fast decoupled variable split.
    Section V-A uses this classical power systems observation to train two separate denoisers, implicitly assuming the cross-couplings are negligible for the generative task.
  • domain assumption The AC power flow equations in eqs. (8)-(9) are the correct physical model for the data generation task.
    Used as the constraint set H(x)=0 for the guidance and as the evaluation metric; this is a standard model in the power systems literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Constrained Diffusion Models for Synthesizing Representative Power Flow Datasets." pith.science (2026). https://pith.science/paper/M4M3S5CR

@misc{pith2026250611281,
  author       = {Pith},
  title        = {Pith review of: Constrained Diffusion Models for Synthesizing Representative Power Flow Datasets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M4M3S5CR}},
  note         = {Machine review of arXiv:2506.11281}
}
read the original abstract

High-quality power flow datasets are essential for training machine learning models in power systems. However, security and privacy concerns restrict access to real-world data, making statistically accurate and physically consistent synthetic datasets a viable alternative. We develop a diffusion model for generating synthetic power flow datasets from real-world power grids that both replicate the statistical properties of the real-world data and ensure AC power flow feasibility. To enforce the constraints, we incorporate gradient guidance based on the power flow constraints to steer diffusion sampling toward feasible samples. For computational efficiency, we further leverage insights from the fast decoupled power flow method and propose a variable decoupling strategy for the training and sampling of the diffusion model. These solutions lead to a physics-informed diffusion model, generating power flow datasets that outperform those from the standard diffusion in terms of feasibility and statistical similarity, as shown in experiments across IEEE benchmark systems.

Figures

Figures reproduced from arXiv: 2506.11281 by the authors.

Figure 1
Figure 1. A high-level view of the diffusion model for synthesizing power [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. illustrates the grid topology and notation used throughout this section. Let B = {1, · · · , B} denote the set of buses and L = {1, · · · , L} denote the set of transmission lines in a power grid. Moreover, let elements of power injection vectors p and q be indexed as pb and qb, and let elements of voltage vectors v and θ be indexed as vb and θb, ∀b ∈ B. In the interest of presentation, we omit shunt admittances in … view at source ↗
Figure 3
Figure 3. Schematic overview of the geometry of sampling (a) without guidance and (b) with manifold-constrained gradient guidance. In sampling without [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Histograms of the ground truth versus synthetic power flow data points for active power injections (first row), reactive power injections (second row), [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: 2D scatter plots with density estimates of the active and reactive power injection (top row), and voltage magnitude and phase angle (bottom row) at [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Histograms of violation magnitudes for the active (top row) and reactive (bottom row) power balance constraints in the PJM 5-bus system, comparing [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Comparison of per-bus (a) mean (MW) and (b) variance (MW2 ) of the active power mismatches, and (c) mean (MVar) and (d) variance (MVar2 ) of the reactive power mismatches for the synthesized data on the IEEE 118-bus system under constrained and unconstrained sampling (…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 39 canonical work pages

  1. [21]

    Domain-constrained diffusion models to synthesize tabular data: A case study in power systems,

    M. Hoseinpour and V . Dvorkin, “Domain-constrained diffusion models to synthesize tabular data: A case study in power systems,” arXiv preprint arXiv:2506.11281, 2025

    arXiv 2025

  2. [40]

    Improving diffusion models for inverse problems using manifold constraints,

    H. Chung et al. , “Improving diffusion models for inverse problems using manifold constraints,” Adv. Neural Inf. Process. Syst. , vol. 35, pp. 25 683–25 696, 2022

    work page 2022

  3. [1]

    PGLearn–an open-source learning toolkit for optimal power flow,

    M. Klamkin et al., “PGLearn–an open-source learning toolkit for optimal power flow,” arXiv preprint arXiv:2505.22825 , 2025

    arXiv 2025

  4. [2]

    A large synthetic dataset for machine learning applications in power transmission grids,

    M. Gillioz et al. , “A large synthetic dataset for machine learning applications in power transmission grids,” Scientific Data, vol. 12, no. 1, p. 168, 2025. 10 TABLE III MEAN AND STANDARD DEVIATION OF TOTAL ACTIVE AND REACTIVE POWER MISMATCHES ACROSS ALL BUSES FOR MODELS TRAINED ON GROUND TRUTH , CONSTRAINED SYNTHETIC , AND UNCONSTRAINED SYNTHETIC DATASE...

    work page 2025

  5. [3]

    Efficient creation of datasets for data-driven power system applications,

    A. Venzke et al. , “Efficient creation of datasets for data-driven power system applications,” Electr. Power Syst. Res, vol. 190, p. 106614, 2021

    work page 2021

  6. [4]

    Machine learning for optimal power flows,

    P. Van Hentenryck, “Machine learning for optimal power flows,” Tu- torials in Operations Research: Emerging Optimization Methods and Modeling Techniques with Applications , pp. 62–82, 2021

    work page 2021

  7. [5]

    Embedding power flow into ma- chine learning for parameter and state estimation,

    L. Pagnier and M. Chertkov, “Embedding power flow into ma- chine learning for parameter and state estimation,” arXiv preprint arXiv:2103.14251, 2021

    arXiv 2021

  8. [6]

    OPF-Learn: An open-source framework for creating representative AC optimal power flow datasets,

    T. Joswig-Jones et al. , “OPF-Learn: An open-source framework for creating representative AC optimal power flow datasets,” 2022 IEEE PES ISGT, pp. 1–5, 2022

    work page 2022

Show all 47 references
  1. [7]

    Generating quality datasets for real-time security assessment: Balancing historically relevant and rare feasible operating conditions,

    A.-A. B. Bugaje et al., “Generating quality datasets for real-time security assessment: Balancing historically relevant and rare feasible operating conditions,” IJEPES, vol. 154, p. 109427, 2023

    2023

  2. [8]

    Privacy-preserving and approximately truthful local electricity markets: A differentially private VCG mechanism,

    M. Hoseinpour et al. , “Privacy-preserving and approximately truthful local electricity markets: A differentially private VCG mechanism,” IEEE Trans. on Smart Grid , vol. 15, no. 2, pp. 1991–2003, 2023

    1991

  3. [9]

    Differentially private algorithms for synthetic power system datasets,

    V . Dvorkin and A. Botterud, “Differentially private algorithms for synthetic power system datasets,” IEEE Control Systems Letters , vol. 7, pp. 2053–2058, 2023

    2023

  4. [10]

    Synthesizing grid data with cyber resilience and privacy guarantees,

    S. Wu and V . Dvorkin, “Synthesizing grid data with cyber resilience and privacy guarantees,” IEEE Control Systems Letters , 2025

    2025

  5. [11]

    Generating synthetic net load data with physics- informed diffusion model,

    S. Zhang et al. , “Generating synthetic net load data with physics- informed diffusion model,” arXiv preprint arXiv:2406.01913 , 2024

    2024 arXiv

  6. [12]

    Generating multivariate load states using a conditional variational autoencoder,

    C. Wang et al., “Generating multivariate load states using a conditional variational autoencoder,” Electr. Power Syst. Res. , vol. 213, p. 108603, 2022

    2022

  7. [13]

    Data-driven EV load profiles generation using a variational auto-encoder,

    Z. Pan et al., “Data-driven EV load profiles generation using a variational auto-encoder,” Energies, vol. 12, no. 5, p. 849, 2019

    2019

  8. [14]

    Generating contextual load profiles using a conditional variational autoencoder,

    C. Wang et al., “Generating contextual load profiles using a conditional variational autoencoder,” in 2022 IEEE PES ISGT-Europe. IEEE, 2022, p. 1–6

    2022

  9. [15]

    Autoencoding beyond pixels using a learned similarity metric,

    A. B. L. Larsen et al. , “Autoencoding beyond pixels using a learned similarity metric,” in ICML. PMLR, 2016, pp. 1558–1566

    2016

  10. [16]

    Generative adversarial network for synthetic time series data generation in smart grids,

    C. Zhang et al., “Generative adversarial network for synthetic time series data generation in smart grids,” in 2018 IEEE SmartGridComm. IEEE, 2018, pp. 1–6

    2018

  11. [17]

    GAN-based model for residential load generation consid- ering typical consumption patterns,

    Y . Gu et al., “GAN-based model for residential load generation consid- ering typical consumption patterns,” in 2019 IEEE PES ISGT . IEEE, 2019, pp. 1–5

    2019

  12. [18]

    Generating energy data for machine learning with recurrent generative adversarial networks,

    M. N. Fekri et al., “Generating energy data for machine learning with recurrent generative adversarial networks,” Energies, vol. 13, no. 1, p. 130, 2019

    2019

  13. [19]

    A data-driven approach for generating synthetic load patterns and usage habits,

    S. El Kababji and P. Srikantha, “A data-driven approach for generating synthetic load patterns and usage habits,” IEEE Trans. Smart Grid , vol. 11, no. 6, pp. 4984–4995, 2020

    2020

  14. [20]

    Model-free renewable scenario generation using gener- ative adversarial networks,

    Y . Chen et al., “Model-free renewable scenario generation using gener- ative adversarial networks,” IEEE Trans. Power Syst., vol. 33, no. 3, pp. 3265–3275, 2018

    2018

  15. [22]

    Learning to solve the AC-OPF using sensitivity- informed deep neural networks,

    M. K. Singh et al. , “Learning to solve the AC-OPF using sensitivity- informed deep neural networks,” IEEE Trans. Power Syst., vol. 37, no. 4, pp. 2833–2846, 2021

    2021

  16. [23]

    Predicting AC optimal power flows: Combining deep learning and lagrangian dual methods,

    F. Fioretto et al., “Predicting AC optimal power flows: Combining deep learning and lagrangian dual methods,” Proc. AAAI Conf. Artif. Intell. , vol. 34, no. 01, pp. 630–637, 2020

    2020

  17. [24]

    Learning optimal solutions for extremely fast AC optimal power flow,

    A. S. Zamzam and K. Baker, “Learning optimal solutions for extremely fast AC optimal power flow,” in 2020 IEEE SmartGridComm . IEEE, 2020, pp. 1–6

    2020

  18. [25]

    Scalable bilevel optimization for gener- ating maximally representative OPF datasets,

    I. V . Nadal and S. Chevalier, “Scalable bilevel optimization for gener- ating maximally representative OPF datasets,” in 2023 IEEE PES ISGT EUROPE. IEEE, 2023, pp. 1–6

    2023

  19. [26]

    On the robustness of machine-learnt proxies for security constrained optimal power flow solvers,

    N. Popli et al., “On the robustness of machine-learnt proxies for security constrained optimal power flow solvers,” Sustain. Energy Grids Netw. , vol. 37, p. 101265, 2024

    2024

  20. [27]

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

    S. Lovett et al., “OPFData: Large-scale datasets for AC optimal power flow with topological perturbations,” arXiv preprint arXiv:2406.07234 , 2024

    2024 arXiv

  21. [28]

    Decision theoretic foundations for conformal predic- tion: Optimal uncertainty quantification for risk-averse agents,

    S. Kiyani et al., “Decision theoretic foundations for conformal predic- tion: Optimal uncertainty quantification for risk-averse agents,” arXiv preprint arXiv:2502.02561, 2025

    2025 arXiv

  22. [29]

    A tutorial on conformal prediction

    G. Shafer and V . V ovk, “A tutorial on conformal prediction.” Journal of Machine Learning Research , vol. 9, no. 3, 2008

    2008

  23. [30]

    Representation learning: A review and new perspec- tives,

    Y . Bengio et al., “Representation learning: A review and new perspec- tives,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 35, no. 8, pp. 1798– 1828, 2013

    2013

  24. [31]

    Closing the loop: A framework for trustworthy machine learning in power systems,

    J. Stiasny et al., “Closing the loop: A framework for trustworthy machine learning in power systems,” arXiv preprint arXiv:2203.07505 , 2022

    2022 arXiv

  25. [32]

    A survey on generative adversarial networks: Variants, applications, and training,

    A. Jabbar et al., “A survey on generative adversarial networks: Variants, applications, and training,” ACM CSUR, vol. 54, no. 8, pp. 1–49, 2021

    2021

  26. [33]

    Diffusion models beat GANs on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat GANs on image synthesis,” Adv. Neural Inf. Process. Syst., vol. 34, pp. 8780–8794, 2021

    2021

  27. [34]

    Short-term wind power scenario generation based on conditional latent diffusion models,

    X. Dong et al. , “Short-term wind power scenario generation based on conditional latent diffusion models,” IEEE Trans. Sustain. Energy , vol. 15, no. 2, pp. 1074–1085, 2023

    2023

  28. [35]

    Diffcharge: Generating EV charging scenarios via a denoising diffusion model,

    S. Li et al. , “Diffcharge: Generating EV charging scenarios via a denoising diffusion model,” IEEE Trans. Smart Grid , vol. 15, no. 4, pp. 3936–3949, 2024

    2024

  29. [36]

    Denoising diffusion probabilistic models,

    J. Ho et al., “Denoising diffusion probabilistic models,” Adv. Neural Inf. Process. Syst., vol. 33, pp. 6840–6851, 2020

    2020

  30. [37]

    A survey of relaxations and approximations of the power flow equations,

    D. K. Molzahn et al. , “A survey of relaxations and approximations of the power flow equations,” Found. Trends Electr. Energy Syst. , vol. 4, no. 1-2, pp. 1–221, 2019

    2019

  31. [38]

    Neural approximate mirror maps for constrained diffusion models,

    B. T. Feng et al. , “Neural approximate mirror maps for constrained diffusion models,” arXiv preprint arXiv:2406.12816 , 2024

    2024 arXiv

  32. [39]

    Consistent diffusion models: Mitigating sampling drift by learning to be consistent,

    G. Daras et al., “Consistent diffusion models: Mitigating sampling drift by learning to be consistent,” Adv. Neural Inf. Process. Syst. , vol. 36, pp. 42 038–42 063, 2023

    2023

  33. [41]

    Boumal, An introduction to optimization on smooth manifolds

    N. Boumal, An introduction to optimization on smooth manifolds . Cambridge University Press, 2023

    2023

  34. [42]

    Supplemental materials,

    M. Hoseinpour and V . Dvorkin, “Supplemental materials,” Available online: https://github.com/miladpourv/constrained-diffusion-powerflow/, 2025

    2025

  35. [43]

    Fast-decoupled power flow method for inte- grated analysis of transmission and distribution systems,

    R. K. Portelinha et al. , “Fast-decoupled power flow method for inte- grated analysis of transmission and distribution systems,” Electr. Power Syst. Res., vol. 196, p. 107215, 2021

    2021

  36. [44]

    Monticelli, State estimation in electric power systems: a generalized approach

    A. Monticelli, State estimation in electric power systems: a generalized approach. Springer Science & Business Media, 2012

    2012

  37. [45]

    J. D. D. Glover and M. S. Sarma, Power system analysis and design . Brooks/Cole Publishing Co., 2001

    2001

  38. [46]

    The power grid library for benchmarking AC optimal power flow algorithms,

    S. Babaeinejad et al. , “The power grid library for benchmarking AC optimal power flow algorithms,” arXiv preprint arXiv:1908.02788, 2019

    1908 arXiv

  39. [47]

    Computational optimal transport,

    G. Peyr ´e and M. Cuturi, “Computational optimal transport,” 2020

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.