Pith. sign in

REVIEW 3 major objections 6 minor 25 references

Restoring Missing Power Grid Data by Pairing Generative AI With Physics Laws

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

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

T0 review · glm-5.2

2026-07-09 16:07 UTC pith:Z6MIPM4R

load-bearing objection Two-stage flow-matching + physics-guided refiner for power system data imputation; architecturally novel but evaluated only on synthetic data generated by the same physics it encodes. the 3 major comments →

arxiv 2607.07250 v1 pith:Z6MIPM4R submitted 2026-07-08 eess.SP

Flow-PIN: A Two-Stage Power-Flow-Guided Method for System-Wide Multivariate Profile Inpainting in Distribution Networks

Zhenghao Zhou , Yiyan Li , Yike Guo , Xinyi Ma , Zheng Yan , Mo-Yuen Chow This is my paper
classification eess.SP PACS 88.05.-b88.05.Ec84.70.+p
keywords datadistributionflow-pininpaintingphysicalprofilecorrelationflow
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Deep generative models can fill gaps in power-distribution measurements with data that looks statistically plausible but violates the physical laws governing electricity flow. This paper proposes Flow-PIN, a two-stage framework that first uses a conditional flow matching model—a generative method that learns to transport random noise toward realistic power-system data distributions—to produce candidate values for missing voltage, active power, and reactive power measurements. The generator is conditioned on both the grid's physical topology and statistical correlations between nodes, and a dynamically weighted physical penalty steers the generation toward physically feasible solutions. In the second stage, a topology-aware refiner uses Laplacian positional encoding to embed grid structure into a transformer, then applies alternating current power-flow equations and a differentiable correlation-alignment loss to correct residual numerical deviations while preserving the temporal fluctuations and spatial dependencies that the generative stage captured. The central claim is that decoupling distribution-aware generation from physics-aware refinement yields inpainted data that is simultaneously numerically accurate, temporally realistic, and spatially consistent—overcoming the trade-off where generative models look realistic but lack precision and regression models are precise but over-smoothed.

Core claim

The paper's core discovery is that the tension between realism and accuracy in missing-data restoration can be resolved by a two-stage architecture: a generative flow-matching model first establishes a statistically faithful prior over power-system profiles, and a subsequent physics-guided refiner corrects local numerical deviations using AC power-flow equations and correlation alignment, without destroying the distributional fidelity the generative stage produced. The ablation study reveals a key structural insight: the AC power-flow penalty in the refinement stage does not minimize point-wise error—in fact, removing it slightly improves numerical accuracy—but it is essential for spatially-

What carries the argument

The central mechanism is the interplay between conditional flow matching (which constructs an optimal-transport path from Gaussian noise to the empirical data distribution, guided by a velocity field estimator) and a topology-aware refiner that uses residual correction. The flow-matching stage conditions on a dual graph structure (topological admittance graph and statistical correlation graph) injected via cross-attention into a U-Net, with a dual-level dynamic weighting strategy that scales the physical penalty by time-step squared and adaptively balances gradient magnitudes. The refiner uses Laplacian positional encoding (eigenvectors of the graph Laplacian) to anchor transformer node-att

Load-bearing premise

The evaluation is conducted on a synthetic dataset built from the IEEE 33-bus system with aggregated residential load profiles and simulated distributed energy resources. The framework's performance under real-world conditions—where noise is non-Gaussian, topology changes dynamically, and operational anomalies create distribution shifts not represented in training data—remains untested.

What would settle it

If applying the TPR refiner to a real-world distribution network with topology changes or abnormal operating conditions causes the spatial correlations and temporal fluctuations established by the CFM stage to degrade rather than improve, the central claim that physics-guided refinement preserves distributional fidelity while improving accuracy would be undermined.

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

If this is right

  • If the decoupling principle generalizes, physics-guided refinement stages could be appended to existing generative imputation pipelines in other networked infrastructure (water, gas, traffic) where physical laws are known but data is missing.
  • The finding that AC power-flow constraints trade local accuracy for global distributional fidelity suggests that evaluation metrics for power-system data restoration should jointly assess point-wise error, temporal dynamics, and spatial correlation rather than relying on any single metric.
  • The differentiable correlation-alignment mechanism, which uses Huber loss on Pearson correlation matrices computed via native tensor operations, could be adopted as a plug-in regularizer for any generative model that must preserve spatial structure in its outputs.
  • The dual-graph conditioning approach (hard topological plus soft statistical edges) offers a template for injecting domain structure into flow-matching architectures beyond power systems.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes Flow-PIN, a two-stage framework for multivariate profile inpainting in distribution networks. Stage 1 uses a conditional flow matching (CFM) model conditioned on topological and correlation graphs, with a dynamically weighted linear voltage-drop penalty integrated into the loss. Stage 2 introduces a topology-aware power-flow-guided refiner (TPR) that uses Laplacian positional encoding, AC power-flow mismatch losses, and a differentiable correlation alignment mechanism to correct residual deviations. The method is evaluated on a synthesized IEEE 33-bus dataset with DERs against ten baselines, showing improvements in point-wise accuracy, temporal fluctuation preservation, and spatial correlation fidelity.

Significance. The paper addresses a practically important problem in distribution network data quality. The two-stage decoupling of distribution-aware generation from physics-aware refinement is a reasonable architectural choice, and the integration of Laplacian positional encoding into a transformer-based refiner is a concrete, falsifiable design decision. The differentiable correlation alignment loss (Eq. 32) using a Huber function is a thoughtful construction for maintaining spatial dependencies during refinement. The ablation study (Fig. 6) and the latent-space alignment analysis provide useful diagnostic evidence for the LPE contribution. The comparison against ten baselines including CSDI, BERT-PIN, and Load-PIN is reasonably comprehensive.

major comments (3)
  1. Section III-A-1: The evaluation dataset is synthesized by running Newton-Raphson AC power flow simulation on the IEEE 33-bus system. The ground-truth data therefore satisfies the same AC power flow equations (Eqs. 28–29) that the TPR stage enforces as constraints. This creates a favorable circularity: the physics-guided components are tested under idealized conditions where the physics model is exactly correct. On real distribution networks, where single-phase equivalent AC power flow is an approximation (line parameter errors, three-phase unbalance, regulator taps, topology uncertainty), these constraints could introduce systematic bias. The authors should explicitly acknowledge this limitation and, ideally, test robustness by injecting line-parameter perturbations or topology noise into the evaluation. This is load-bearing because the central claim of physics-guided improvement restson
  2. Fig. 6(b) and accompanying text: The ablation removing the AC-PF penalty (w/o AC-PF) yields lower point-wise numerical errors (NMAE, NRMSE) than the complete model, and the authors acknowledge this trade-off. This result is actually informative: it suggests the AC-PF penalty sacrifices local accuracy for distributional/spatial fidelity. However, under the synthetic evaluation where the physics model is exact, this trade-off is already unfavorable for point-wise metrics. Under model mismatch, the trade-off would likely worsen. The authors should discuss whether the AC-PF penalty weight (λ_pf) was tuned, and whether a different weighting could recover point-wise accuracy without sacrificing spatial fidelity. The current presentation frames this as an acceptable trade-off without quantifying the sensitivity to λ_pf.
  3. The framework involves a substantial number of free parameters (λ_mask, α, λ_v, λ_pf, λ_c, α_v, α_pq, β, λ_reg, k). The manuscript does not report the values used, the tuning procedure, or sensitivity analysis for most of these. Given that the dynamic weighting (Eq. 19) and the refiner loss balancing (Eq. 33) depend on these, reproducibility is compromised. At minimum, the values used in the reported experiments and a brief note on how they were selected should be provided, ideally in an appendix or supplementary material.
minor comments (6)
  1. Eq. (23): The Laplacian matrix L is defined, but the equation as written appears to show L = D^{-1/2}(I - A)D^{-1/2} or similar, which is the normalized Laplacian. The notation should be clarified to distinguish between the combinatorial and normalized Laplacian, since the eigenvector computation in Eq. (24) depends on which is used.
  2. References [9] and [10] list publication years of 2026, which appears to be a typographical error. Please verify and correct.
  3. Table I: The WD values for Flow-PIN on active power (0.0047) and reactive power (0.0016) are substantially lower than the next-best baseline. Given that WD is computed on complete datasets, clarifying whether the imputed segments alone or the full reconstructed datasets are compared would be informative.
  4. Section III-A-1: The dataset description mentions 0.1% Gaussian white noise injection but does not specify whether this noise is added before or after the power flow simulation. This affects interpretation of the physics-consistency claim.
  5. Fig. 2 caption: The figure description is detailed but the text references 'E_graph' and 'E_node' with specific formatting that may not render correctly in all formats. Please verify figure legibility in the final format.
  6. The term 'profile inpainting' is used throughout but is not formally distinguished from 'data imputation' or 'missing data restoration' in the problem formulation (Section II-A). A brief note on terminology would help readers coming from adjacent fields.

Circularity Check

0 steps flagged

No significant circularity; the physics-in-data-generation concern is a generalization risk, not a definitional circularity.

full rationale

The paper's derivation chain is self-contained. Stage 1 (CFM) learns a generative vector field via flow matching with a physics penalty (Eq. 16); Stage 2 (TPR) applies residual corrections guided by AC-PF mismatch (Eqs. 28-29) and correlation alignment (Eq. 32). None of these steps reduce to their inputs by construction: the model's predictions are not defined in terms of the ground truth, and the physics constraints are loss terms, not identities that force the output. The skeptic's concern that the evaluation dataset is generated by AC power flow simulation (the same equations used as constraints) is a legitimate generalization/external-validity risk, but it is not circularity in the definitional sense — the method does not trivially recover the data because the physics constraints hold. Indeed, the ablation in Fig. 6(b) shows that removing the AC-PF penalty yields *lower* point-wise errors, which would be inexplicable if the physics constraints were satisfied by construction. The only minor self-citation is [15] (Load-PIN, co-authored by Yiyan Li), used as a baseline comparison rather than as a load-bearing theoretical premise. This is normal practice and does not raise the circularity score meaningfully.

Axiom & Free-Parameter Ledger

10 free parameters · 4 axioms · 0 invented entities

The paper introduces several hyperparameters and loss weights that are not specified in the text. The core axioms are domain assumptions about the validity of linearized power flow, the representativeness of the synthetic dataset, and the adequacy of Pearson correlation for capturing spatial dependencies. No new physical entities or forces are invented.

free parameters (10)
  • λ_mask
    Hyperparameter scaling the masked loss in the CFM stage (Eq 20). Value not stated.
  • α
    Decay factor for the exponential moving average in the dynamic weighting strategy (Eq 19). Value not stated.
  • λ_v
    Weight for the voltage anchoring penalty in the TPR loss (Eq 33). Value not stated.
  • λ_pf
    Weight for the physical mismatch loss in the TPR loss (Eq 33). Value not stated.
  • λ_c
    Weight for the correlation alignment loss in the TPR loss (Eq 33). Value not stated.
  • α_v
    Weight for voltage in the data fidelity loss (Eq 25). Value not stated.
  • α_pq
    Weight for power in the data fidelity loss (Eq 25). Value not stated.
  • β
    Threshold parameter for the Huber function in the correlation alignment loss (Eq 31). Value not stated.
  • λ_reg
    Regularization factor for phase angles in the physical mismatch loss (Eq 30). Value not stated.
  • k
    Number of smallest non-trivial eigenvalues for Laplacian Positional Encoding (Eq 24). Value not stated.
axioms (4)
  • domain assumption Linearized power flow approximation using nodal impedance matrix is valid for the physical penalty in the generative stage.
    Used in Eq 16 to formulate the physics-guided voltage drop penalty. This is a simplification of AC power flow, assumed sufficient to guide generation without suppressing distribution learning.
  • domain assumption The synthesized IEEE 33-bus dataset with aggregated residential loads and DERs is representative of real active distribution networks.
    The entire evaluation rests on this dataset. The paper states it 'accurately simulate the diversified load characteristics' but this is an unverified assumption regarding real-world complexity.
  • domain assumption Pearson correlation coefficients adequately capture the spatial dependencies needed for the correlation graph and alignment loss.
    Used in Eq 5 and Eq 32. Assumes linear correlation is the primary statistical relationship to preserve, which may not capture all non-linear dependencies.
  • ad hoc to paper The dual-level dynamic weighting strategy for the physical loss prevents gradient conflicts without disrupting the generative prior.
    The design of the temporal coefficient (t^2) and adaptive gradient balancing (Eq 19) is introduced to solve a specific optimization conflict. Its general effectiveness is assumed.

pith-pipeline@v1.1.0-glm · 23450 in / 2287 out tokens · 372853 ms · 2026-07-09T16:07:11.928938+00:00 · methodology

0 comments
read the original abstract

High-quality system measurement data is critical for power distribution system operation. As deep generative models (e.g., GAN, Diffusion, etc.) have been widely studied to solve the missing data restoration problem to enhance the data quality, their results may look "realistic" but not sufficiently "accurate" due to lacking physical guarantees. To address this limitation, a two-stage physics-guided framework, Flow-PIN, is proposed in this paper for system-wide multivariate profile inpainting. The first stage employs a conditional flow matching model, conditioned on topological and correlation graphs, to generate candidate values. A physical penalty is integrated into the loss function to constrain the generative vector field based on grid physical laws. The second stage introduces a topology-aware power-flow-guided refiner that utilizes Laplacian positional encoding to inject topology information into node embeddings. By coupling alternating current power flow equations with a differentiable correlation alignment mechanism, this refiner further corrects numerical deviations. Evaluations on an active distribution network dataset benchmark the proposed framework against ten representative baselines. The results show that Flow-PIN achieves high-fidelity profile inpainting across three dimensions: maximizing numerical accuracy, capturing temporal fluctuation, and preserving spatial topological correlations.

Figures

Figures reproduced from arXiv: 2607.07250 by Mo-Yuen Chow, Xinyi Ma, Yike Guo, Yiyan Li, Zhenghao Zhou, Zheng Yan.

Figure 1
Figure 1. Figure 1: The workflow of data preprocessing. To simulate realistic missing values, a mixed spatio￾temporal mask matrix M ∈ {0,1}T×N×3 is constructed. A matrix element value of 1 indicates an available value and 0 indicates a missing one. The generation process simulates continuous time block failures rather than dropping random isolated points. The observed incomplete input data is defined as: X M X obs   Norm( )… view at source ↗
Figure 2
Figure 2. Figure 2: The framework consists of two stages: the CFM module (top) and the TPR module (bottom). In the CFM stage, Gaussian noise, observed data, and the missingness mask are concatenated as inputs. Historical node features H are extracted from observations and processed by a graph encoder alongside dual-track priors (topological and correlation graphs) to form graph embeddings Egraph. The Egraph is injected into t… view at source ↗
Figure 3
Figure 3. Figure 3: The IEEE 33-bus distribution network [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comprehensive spatiotemporal and physical validation of the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation results of Flow-PIN. The grouped radar charts compare the normalized errors of the complete model and its variants for the CFM stage, the TPR stage, and the complete two-stage framework [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages · 4 internal anchors

  1. [1]

    6 (a) evaluates the components of CFM

    Fig. 6 (a) evaluates the components of CFM. Removing the dual graph conditions increases the APCE, indicating that topological and statistical conditions are necessary for capturing spatial coupling; without them, the generative model processes nodes in isolation. Furthermore, rem oving the physical penalty increases errors across all axes, demonstrating ...

    work page 1962

  2. [2]

    A Data Imputation Model in Phasor Measurement Units Based on Bagged Averaging of Multiple Linear Regression,

    N. T. Le and W. Benjapolakul, “A Data Imputation Model in Phasor Measurement Units Based on Bagged Averaging of Multiple Linear Regression,” IEEE Access, vol. 6, pp. 39324 – 39333, 2018, doi: 10.1109/ACCESS.2018.2856768

    work page doi:10.1109/access.2018.2856768 2018

  3. [3]

    Towards missing electric power data imputation for energy management systems,

    M. - C. Wang, C. - F. Tsai, and W. - C. Lin, “Towards missing electric power data imputation for energy management systems,” Expert Systems with Applications, vol. 174, p. 114743, Jul. 2021, doi: 10.1016/j.eswa.2021.114743

    work page doi:10.1016/j.eswa.2021.114743 2021

  4. [4]

    Robust Recovery of Missing Data in Electricity Distribution Systems,

    C. Genes, I. Esnaola, S. M. P erlaza, L. F. Ochoa, and D. Coca, “Robust Recovery of Missing Data in Electricity Distribution Systems,” IEEE Trans. Smart Grid, vol. 10, no. 4, pp. 4057 – 4067, Jul. 2019, doi: 10.1109/TSG.2018.2848935

    work page doi:10.1109/tsg.2018.2848935 2019

  5. [5]

    A Regularized Tensor Completion Approach for PMU Data Recovery,

    A. Ghasemkhani, I. Niazazari, Y. Liu, H. Livani, V . A. Centeno, and L. Yang, “A Regularized Tensor Completion Approach for PMU Data Recovery,” IEEE Trans. Smart Grid, vol. 12, no. 2, pp. 1519 – 1528, Mar. 2021, doi: 10.1109/TSG.2020.3030566

    work page doi:10.1109/tsg.2020.3030566 2021

  6. [6]

    An Alte rnating Direction Method of Multipliers Based Approach for PMU Data Recovery,

    M. Liao, D. Shi, Z. Yu, Z. Yi, Z. Wang, and Y. Xiang, “An Alte rnating Direction Method of Multipliers Based Approach for PMU Data Recovery,” IEEE Trans. Smart Grid, vol. 10, no. 4, pp. 4554 – 4565, Jul. 2019, doi: 10.1109/TSG.2018.2864176

    work page doi:10.1109/tsg.2018.2864176 2019

  7. [7]

    Electricity Measurement Missing Data Recovery Method Based on Low - Rank Tensor Completion With Triangular Canonical Polyadic Decomposition,

    T. Yang, Z. Yang, and H. Pen, “Electricity Measurement Missing Data Recovery Method Based on Low - Rank Tensor Completion With Triangular Canonical Polyadic Decomposition,” IEEE Trans. Instrum. Meas., vol. 74, pp. 1 – 13, 2025, doi: 10.1109/TIM.2025.3617412

    work page doi:10.1109/tim.2025.3617412 2025

  8. [8]

    Corre lation Clustering Imputation for Diagnosing Attacks and Faults With Missing Power Grid Data,

    R. Razavi - Far, M. Farajzadeh - Zanjani, M. Saif, and S. Chakrabarti, “Corre lation Clustering Imputation for Diagnosing Attacks and Faults With Missing Power Grid Data,” IEEE Trans. Smart Grid, vol. 11, no. 2, pp. 1453 – 1464, Mar. 2020, doi: 10.1109/TSG.2019.2938251

    work page doi:10.1109/tsg.2019.2938251 2020

  9. [9]

    Missing Data Recovery in Large Power Systems Using Network Embedding,

    T. Wu, Y. - J. A. Zhang, Y. Liu, W. C. Lau, and H. Xu, “Missing Data Recovery in Large Power Systems Using Network Embedding,” IEEE Trans. Smart Grid, vol. 12, no. 1, pp. 680 – 691, Jan. 2021, doi: 10.1109/TSG.2020.3014813

    work page doi:10.1109/tsg.2020.3014813 2021

  10. [10]

    Missing Data Recovery for Electric Load Curves Based on A daptive Curve Grafting and Noise Decomposition,

    N. Xu, L. Fu, X. Wang, and Z. Lin, “Missing Data Recovery for Electric Load Curves Based on A daptive Curve Grafting and Noise Decomposition,” IEEE Trans. Power Syst., pp. 1 – 12, 2026, doi: 10.1109/TPWRS.2026.3655678

    work page doi:10.1109/tpwrs.2026.3655678 2026

  11. [11]

    ZTFed - MAS2S: A Zero - Trust Federated Learning Framework With Verifiable Privacy and Trust - Aware Aggregation for Wind Power Data Imputation,

    Y. Li, H. Wang, Y. Li, J. Li, and Z. Dong, “ZTFed - MAS2S: A Zero - Trust Federated Learning Framework With Verifiable Privacy and Trust - Aware Aggregation for Wind Power Data Imputation,” IEEE Trans. Ind. Inf., vol. 22, no. 1, pp. 165 – 175, Jan. 2026, doi: 10.1109/TII.2025.3609075

    work page doi:10.1109/tii.2025.3609075 2026

  12. [12]

    Unsupervised and controllable synthesizing f or imbalanced energy dataset based on AC - InfoGAN,

    Z. Zhou, Y. Li, R. Liu, X. Xu, and Z. Yan, “Unsupervised and controllable synthesizing f or imbalanced energy dataset based on AC - InfoGAN,” Applied Energy, vol. 393, p. 126107, Sep. 2025, doi: 10.1016/j.apenergy.2025.126107

    work page doi:10.1016/j.apenergy.2025.126107 2025

  13. [13]

    ProfileSR - GAN: A GAN Based Super - Resolution Method for Generating High - Resolution Load Profiles,

    L. Song, Y. Li, and N. Lu, “ProfileSR - GAN: A GAN Based Super - Resolution Method for Generating High - Resolution Load Profiles,” IEEE Trans. Smart Grid, vol. 13, no. 4, pp. 3278 – 3289, Jul. 2022, doi: 10.1109/tsg.2022.3158235

    work page doi:10.1109/tsg.2022.3158235 2022

  14. [14]

    Multivariate time series imputation with generative adversarial networks,

    Y. Luo, X. Cai, Y. Zhang, and J. Xu, “Multivariate time series imputation with generative adversarial networks,” Advances in neural information processing systems, vol. 31, 2018

    work page 2018

  15. [15]

    A Novel Method for Missing Data Reconstruction in Smart Grid Using Generative Adversarial Networks,

    J. Fang, L. Zheng, and C. Liu, “A Novel Method for Missing Data Reconstruction in Smart Grid Using Generative Adversarial Networks,” IEEE Trans. Ind. Inf., vol. 20, no. 3, pp. 4408 – 4417, Mar. 2024, doi: 10.1109/TII. 2023.3306366

    work page doi:10.1109/tii 2024

  16. [16]

    Load Profile Inpainting for Missing Load Data Restoration and Baseline Estimation,

    Y. Li et al., “Load Profile Inpainting for Missing Load Data Restoration and Baseline Estimation,” IEEE Trans. Smart Grid, vol. 15, no. 2, pp. 2251 – 2260, Mar. 2024, doi: 10.1109/TSG.2023.3293188

    work page doi:10.1109/tsg.2023.3293188 2024

  17. [17]

    A non - intrusive load monitoring missing data repair method for power grids based on an improved state transfer algorithm,

    W. Sun, Y. Hao, B. Gao, L. Zhao, T. Wang, and L. Xing, “A non - intrusive load monitoring missing data repair method for power grids based on an improved state transfer algorithm,” Electric Power Systems Research, vol. 248, p. 111874, Nov. 2025, doi: 10.1016/j.epsr.2025.111874

    work page doi:10.1016/j.epsr.2025.111874 2025

  18. [18]

    A Cross - level Consistency - aware Diffusion Model of Hierarchical Imputation for Energy Prosumers,

    H. Dong, J . Zhu, C. Chung, Y. Chen, H. Yang, and Z. Liang, “A Cross - level Consistency - aware Diffusion Model of Hierarchical Imputation for Energy Prosumers,” IEEE Trans. Smart Grid, pp. 1 – 1, 2025, doi: 10.1109/TSG.2025.3611142

    work page doi:10.1109/tsg.2025.3611142 2025

  19. [19]

    BERT - PIN: A BERT - Based Framework for Recovering Missing Data Segments in Time - Series Load Profiles,

    Y. Hu, K. Ye, H. Kim, and N. Lu, “BERT - PIN: A BERT - Based Framework for Recovering Missing Data Segments in Time - Series Load Profiles,” IEEE Trans. Ind. Inf., vol. 20, no. 10, pp. 12241 – 12251, Oct. 2024, doi: 10.1109/TII.2024.3417272

    work page doi:10.1109/tii.2024.3417272 2024

  20. [20]

    Applying fine - tu ned LLMs for reducing data needs in load profile analysis,

    Y. Hu, H. Kim, K. Ye, and N. Lu, “Applying fine - tu ned LLMs for reducing data needs in load profile analysis,” Applied Energy, vol. 377, p. 124666, Jan. 2025, doi: 10.1016/j.apenergy.2024.124666

    work page doi:10.1016/j.apenergy.2024.124666 2025

  21. [21]

    A Large - Scale Dataset of Distributed Renewable Energy Scenarios on the IEEE - 33 Bus Network,

    Y. Chen, H. Xie, W. Huang, and P. Li, “A Large - Scale Dataset of Distributed Renewable Energy Scenarios on the IEEE - 33 Bus Network,” Sci Data, vol. 13, no. 1, p. 146, Dec. 2025, doi: 10.1038/s41597 - 025 - 06464 - w

    work page doi:10.1038/s41597 2025

  22. [22]

    A Time Series is Worth 64 Words: Long-term Forecasting with Transformers

    Y. Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A T ime Series is Worth 64 Words: Long - term Forecasting with Transformers,” Mar. 05, 2023, arXiv: arXiv:2211.14730. doi: 10.48550/arXiv.2211.14730

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2211.14730 2023

  23. [23]

    Are Transformers Effective for Time Series Forecasting?

    A. Zeng, M. Chen, L. Zhang, and Q. Xu, “Are Transformers Effective for Time Series Forecasting?,” Aug. 17, 2022, arXiv: arXiv:2205.13504. doi: 10.48550/arXiv.2205.13504

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2205.13504 2022

  24. [24]

    TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis

    H. Wu, T. Hu, Y. Liu, H. Zhou, J. Wang, and M. Long, “TimesNet: Temporal 2D - Variation Modeling for General Time Series Analysis,” Apr. 11, 2023, arXiv: arXiv:2210.02186

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  25. [25]

    CSDI: Conditional Score-based Diffusion Models for Probabilistic Time Series Imputation

    Y. Tashiro, J . Song, Y. Song, and S. Ermon, “CSDI: Conditional Score - based Diffusion Models for Probabilistic Time Series Imputation,” Oct. 27, 2021, arXiv: arXiv:2107.03502. doi: 10.48550/arXiv.2107.03502

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2107.03502 2021