Pith. sign in

REVIEW 4 major objections 4 minor 58 references

SynEnergy: Anomaly Semantic-Guided Diffusion for Synthetic Energy Data Generation

T0 review · 4 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read SynEnergy claims that explicitly modeling regional anomaly semantics and injecting them into a diffusion generator preserves rare energy events in synthetic data, reporting a 12.21% average anomaly-fidelity improvement over 11 baselines.

desk verdict Real framework, solid engineering, but the headline anomaly-fidelity gain is likely inflated because the anomaly labels and the evaluation metrics share the same residual threshold the model is trained to match. read the letter →

arxiv 2608.03087 v1 pith:ZSHZ2AOO submitted 2026-08-04 cs.LG cs.AI

Lin Jiang , Dahai Yu , Ravikumar Gelli , Guang Wang This is my paper · ORCID
classification cs.LGcs.AI
keywords energyconsumptiondatasyntheticgenerationdiffusionmodelsanomalypreservationheterogeneousgraphssemanticstimeseriesdownstreamutility
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

The paper tries to establish that synthetic energy consumption data can faithfully include rare anomalous events—outages, heatwaves, behavioral shifts—rather than smoothing them away, and that the key to doing so is making the anomaly itself a first-class learned object. It proposes SynEnergy, a two-stage diffusion framework: the first stage learns region-specific anomaly semantics from sparse residual patterns while modeling dependencies between regions through spatial and attribute graphs; the second stage injects those semantics into the denoising process through layer-wise control signals. The authors report that on four real-world datasets, against 11 general-purpose and energy-specific baselines, SynEnergy improves anomaly preservation fidelity by an average of 12.21% and downstream quality by 2.96% while keeping overall generation fidelity competitive. A sympathetic reader would care because downstream applications such as demand response planning and grid reliability assessment depend on the rare events that standard generators underrepresent.

What carries the argument

The load-bearing object is the anomaly semantic: a centroid of household residual embeddings within a region, after cosine-similarity pooling, that has been enriched by a graph attention layer using spatial and attribute adjacency matrices as structural priors. This object carries the argument because it converts sparse, localized deviations into a compact, sampleable condition that the diffusion decoder can be guided by. On the generation side, the control module turns the semantic into global and temporal anomaly hints, modulates them by region through adaptive layer normalization, and injects layer-wise control features into the backbone denoiser's encoder states, gated both by the predicted anomaly occurrence probability and by a diffusion-step schedule. The backbone denoiser is a Transformer that splits its estimate into trend and seasonal components, so the regular consumption pattern and the anomaly pattern are handled by separate mechanisms.

What would settle it

Run SynEnergy on a dataset with independently verified event logs (outage tickets, landfall timings, heat advisories) and compare generated sequences against those logs using event-level metrics rather than residual thresholds; if synthetic data matches residual statistics but misses logged events, the anomaly semantic is optimizing the wrong target. A simpler variant: swap in a different anomaly detector, such as change-point detection, and check whether the reported 12.21% improvement survives.

Watch

Extended reading notes

Core claim

The central claim is that anomalies in energy data are not independent noise: they cluster geographically and along socioeconomic attributes, and modeling those dependencies explicitly is what allows a generative model to reproduce them. Concretely, SynEnergy defines an anomaly as a residual from a region-wide background consumption sequence, builds household-level residual embeddings, pools similar embeddings into regional anomaly spaces, enhances those spaces with gated cross-region attention over spatial and attribute adjacency graphs, and samples a graph-enhanced semantic to condition reverse diffusion. The generated sequences then reproduce not only the overall distribution but also event timing, magnitude, and household impact—for example, roughly 87% of sampled households show prolonged zero-consumption during the Hurricane Michael period, close to the original data. On the main Florida hurricane dataset, SynEnergy ranks first on all four anomaly-fidelity metrics and both downstream utility metrics, with the reported 12.21% average anomaly-fidelity improvement over the strongest baseline.

Load-bearing premise

Everything rests on defining an anomaly as a residual from the region-wide average consumption, then keeping the top 10% of absolute residuals; if that regional baseline or percentile choice does not match what actually counts as an anomalous event, the model learns and preserves the wrong events.

Editorial extensions

If this is right

  • Synthetic data from SynEnergy can train downstream anomaly-detection and prediction models: under the train-on-synthetic, test-on-real protocol, both detection and prediction PR-AUC improve by an average of 2.96% over the strongest baselines.
  • Generation remains competitive or best from 100 to 50,000 households and across 1-hour, 4-hour, and 1-day temporal granularities, so the method scales to city-wide settings.
  • City-scale coordinated anomalies—hurricane-driven outages and heatwave demand surges—are reproduced with their timing and household-level reach, not just their distributional footprint.
  • The anomaly threshold is not neutral: the 90th-percentile default works best, while very restrictive (1%) and very broad (20%) thresholds degrade results, so the method is tuned to a moderate definition of 'anomalous'.

Reading between the lines

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

  • The paper measures anomalies only through its residual-threshold definition; an immediate extension would be to validate against independently logged events such as outage tickets and weather advisories, which would show whether the preserved anomalies coincide with operationally meaningful ones.
  • Because an anomaly semantic is a sampleable object, the framework could in principle support controllable anomaly-rate generation—producing datasets with more or fewer rare events—but the paper does not explore this dial.
  • The same two-stage recipe of residual semantics, graph enhancement, and control injection may transfer to other spatiotemporal domains with sparse, region-correlated events, such as traffic incidents or water-network failures; this is an extrapolation, not a paper claim.
  • Even SynEnergy under-generates the rarest events: its zero-consumption anomaly rate is 7.94% versus 9.14% in real data, so closing that remaining gap is a natural next step.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper addresses the problem of preserving anomalous events when generating synthetic energy consumption data. It defines anomalies as intervals whose deviation from a regional background sequence exceeds a percentile threshold, and proposes SynEnergy, a two-stage framework: HG-ASL learns region-specific anomaly semantics from filtered residuals using clustering and a heterogeneous graph over spatial and attribute similarities, while AS-Diff injects the sampled semantics into a Transformer-based diffusion denoiser through layer-wise control. The authors evaluate on four datasets (FL1, FL2, NY, CA) against 11 baselines, reporting an average improvement of 12.21% in anomaly preservation fidelity and 2.96% in downstream quality while maintaining competitive overall fidelity. The appendix provides extensive additional tables and figures.

Significance. If the reported results hold, the paper makes a useful contribution by explicitly targeting anomaly preservation in synthetic energy data generation and by demonstrating a two-stage conditioning mechanism that couples residual-based anomaly semantics with diffusion generation. The empirical scope is substantial: four datasets, 11 baselines, multiple fidelity and utility metrics, and a detailed appendix. The idea of learning anomaly semantics from sparse residuals and injecting them into the denoising process is likely to be of interest to the time-series generation community. However, the current significance is tempered by concerns about the consistency and validity of the anomaly evaluation metrics, which are discussed below.

major comments (4)
  1. [Section 3.2 / Eq. (36)] The definition of the residual baseline is inconsistent between real and generated data, which directly affects all anomaly-fidelity metrics. In Section 3.2, b_n is described as an expected regional pattern estimated from historical consumption records, whereas Eq. (36) computes the generated residual as \bar{x} - \bar{b}_n, where \bar{b}_n is the regional background of the generated data. If b_n is a fixed historical profile and \bar{b}_n is the average of the generated sequences, the two residual distributions are not comparable: a generator that correctly reproduces a city-wide outage (Section 5.4 reports about 87% of households affected) will have a depressed \bar{b}_n during the outage, so \bar{e} will be small and A-Rate will not credit the correct reproduction. If instead b_n is the contemporaneous cross-household mean, then the anomaly definition cannot label widespread events at all, because the background itself drops when most households are affected. Please clarify which baseline is used and recompute the anomaly metrics with a single fixed reference profile for both real and generated data.
  2. [Section 3.2 / Appendix D.3] The anomaly threshold \delta is reused across training, evaluation, and downstream labeling. Section 4.1.1 filters residuals with soft weights derived from the same residuals, and Eqs. (37), (52), and (55) define all anomaly-fidelity and downstream labels using the same \delta. Since no baseline is trained with this threshold, the reported 12.21% improvement may measure alignment with the chosen evaluation criterion rather than preservation of externally meaningful anomalies. Table 9 and Appendix D.7 show sensitivity to \delta, and the authors acknowledge that 1% and 20% thresholds degrade performance. Please add an evaluation using an independent anomaly definition (e.g., known event windows or an unsupervised detector not based on the same residuals) and report the headline gains as a function of \delta.
  3. [Section 4.2] The training objective for AS-Diff is not specified. The section describes the Backbone Denoiser and AS-Control architecture and the reverse update in Eq. (11), but does not state the loss function used to optimize the denoiser and control network, nor the values of \gamma_e, \tau_e, \lambda_reg, \lambda_pre, and \lambda_ctrl introduced in Eqs. (2), (10), and (16). The Introduction refers to a pretrained Backbone Denoiser, but the pretraining and fine-tuning protocol is not described. Please add the complete training procedure and hyperparameter settings so that the central method is reproducible from the manuscript.
  4. [Tables 1 and 4-6] The headline '12.21%' and '2.96%' averages are not defined precisely. On FL1, A-Tail ties with HeavyDiff, yielding 0% improvement on that metric; in Table 4 SynEnergy's Pred-PRAUC is lower than Diffusion-TS (0.2411 vs 0.2421); and in Table 5 Pred-PRAUC ties with Diffusion-TS at 0.6109. Please report per-dataset, per-metric comparisons against the strongest baseline, define the averaging procedure, and state whether any differences are statistically significant.
minor comments (4)
  1. [Section 5.4 / Figure 10] The claim that about 87% of sampled households exhibit prolonged zero-consumption anomalies is not visible in the figure, which appears to show only mean consumption; please provide the supporting per-household statistic or a reference to a table that quantifies it.
  2. [Appendix D.2] Reference [9] is cited as the source of the spectrogram-conditioning adaptation for DiffWave, but [9] is about model extraction defense and seems unrelated; please verify the citation.
  3. [Abstract / Section 5.2] The code availability statement points to 'SynEnergy Code' without a working URL; please provide a repository link or a clear statement of how the code can be obtained.
  4. [Sections 4.1 and 5.1.4] The hyperparameters \gamma_e and \tau_e from Eq. (2) are not given in Section 5.1.4 or in the appendix; please report all hyperparameter values in a single table.

Circularity Check

1 steps flagged · score 2.0 of 10

Mild self-referential anomaly definition: the residual-threshold events the model is built to reproduce also define every anomaly-fidelity and downstream metric; external baselines and data-based metrics keep the central comparison non-circular.

  1. self definitional [Appendix D.3 (Metric Description); see also Section 3.2 and Eqs. (36)-(37), (52), (55)]
    "we identify anomalous points using the residual-based formulation introduced in Section 3.2. ... An interval is identified as anomalous when the absolute residual exceeds the anomaly threshold δ: a^k_{n,m} = I(|e^k_{n,m}| > δ) ... (Eq. 37). The detection label at interval k is defined as ... I(|e^k_{n,m}| > δ ∧ |e^{k-1}_{n,m}| ≤ δ) (Eq. 52)."

    The pipeline is closed around one residual-threshold definition. Section 3.2 defines anomalies by thresholding residuals e = x - b with δ set to a percentile of the same residual distribution. HG-ASL learns its anomaly semantics from these residual sequences, and AS-Diff injects them into generation. All four anomaly-preservation metrics (A-Rate, A-Count, A-Energy, A-Tail) and the downstream labels in Eqs. (52) and (55) then re-threshold residuals with the same δ. Thus the headline 12.21% improvement measures how well the model reproduces the exact residual-exceedance patterns it was built to represent; the evaluation target is, by construction, the model's own anomaly representation. The choice of δ is also tuned in Appendix D.7 and Table 9 partly because it yields the best metrics.

full rationale

The central derivation chain is otherwise self-contained: SynEnergy is trained on real consumption data and evaluated against 11 external baselines with metrics computed from generated data, not from the model's fitting losses. The backbone denoiser follows Diffusion-TS, an external method, and the anomaly semantics are clustered residual embeddings rather than a parameter fitted to the evaluation metric. There is no load-bearing self-citation, no imported uniqueness theorem, and no renamed known result: the graph-based semantic learning and control injection are original mechanisms. The only circularity concern is the self-referential anomaly definition: the same residual threshold δ that defines the anomaly semantics in Section 3.2 and HG-ASL also defines every anomaly-preservation metric and downstream label. This does not force the reported 12.21% improvement mathematically, because the generated sequences are stochastic and the baseline comparisons are data-based, but it does mean the headline number evaluates self-consistency with the paper's own residual-threshold notion of anomaly rather than an independently grounded event definition. Accordingly, the score is 2: one mild self-referential evaluation loop with substantive independent content elsewhere.

Assumptions & free parameters 4 free parameters · 3 assumptions · 1 invented entities

The central claim depends on several fitted hyperparameters, especially the anomaly threshold delta, and on the domain assumption that residual deviations from a regional average capture meaningful anomalies. The anomaly semantics themselves are internal representations, not independently observable entities.

free parameters (4)
  • anomaly threshold delta = 90th percentile of absolute residuals on each dataset
    Defines what counts as an anomaly in Section 3.2; all anomaly semantics, metrics, and downstream labels depend on it. Table 9 shows performance varies across top 1%, 5%, 10%, and 20% settings.
  • soft-weighting parameters gamma_e and tau_e = not specified numerically
    Control the soft weighting of residuals in Eq. (2); no sensitivity analysis or stated values in the main text.
  • diffusion-step gate lambda_ctrl = not specified
    Controls injection strength of anomaly guidance over diffusion steps in Eq. (16); no analysis of its effect is reported.
  • loss weights lambda_reg and lambda_pre = not specified
    Balance global, regional, and preservation losses in Eq. (10); no ablation or stated values.
assumptions (3)
  • domain assumption The regional background sequence b_n, computed by averaging consumption across households in a region, is an adequate baseline for expected consumption.
    Section 3.2 defines residuals by subtracting this background; if the background is biased, the anomaly labels are biased.
  • domain assumption Geographical proximity and attribute similarity graphs capture the main cross-region dependencies of energy anomalies.
    The motivational analysis in Section 2 shows correlations, but the graph construction and its completeness are not validated.
  • domain assumption Diffusion denoisers with trend-seasonality decomposition can model regular consumption patterns well enough for the anomaly injection to be beneficial.
    The Backbone Denoiser follows Diffusion-TS; if the backbone itself distorts anomalies, the control injection may only partially compensate.
invented entities (1)
  • anomaly semantics u and graph-enhanced anomaly semantics u_bar
    purpose: Latent representations of regional anomalous patterns used to guide diffusion generation.
    These are abstract learned representations without a falsifiable handle outside the paper; their utility is only shown through downstream metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SynEnergy: Anomaly Semantic-Guided Diffusion for Synthetic Energy Data Generation." pith.science (2026). https://pith.science/paper/ZSHZ2AOO

@misc{pith2026260803087,
  author       = {Pith},
  title        = {Pith review of: SynEnergy: Anomaly Semantic-Guided Diffusion for Synthetic Energy Data Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZSHZ2AOO}},
  note         = {Machine review of arXiv:2608.03087}
}
read the original abstract

Fine-grained energy consumption data are essential for applications such as demand forecasting, demand response planning, and grid reliability assessment. However, access to such data is often restricted by privacy concerns and data-sharing constraints, motivating growing interest in synthetic energy data generation. Although existing methods can reproduce overall consumption distributions and recurring temporal patterns, they often smooth out or underrepresent anomalous events caused by extreme weather, infrastructure failures, and behavioral shifts. Preserving these events is challenging because they are sparse, localized in time and space, and shaped by heterogeneous dependencies across geographical proximity and regional attributes. To address these challenges, we propose SynEnergy, a two-stage diffusion-based framework for anomaly-preserving energy consumption data generation. The first stage, Heterogeneous Graph-based Anomaly Semantic Learning (HG-ASL), extracts region-specific anomaly semantics from sparse residual structures by jointly modeling spatial and attribute dependencies across urban regions. The second stage, Anomaly Semantic-guided Diffusion (AS-Diff), injects the learned anomaly semantics into the denoising process to generate realistic consumption sequences while preserving anomalous patterns. This design enables controllable generation for individual regions and scales naturally to city-wide settings. We evaluate SynEnergy on four real-world energy consumption datasets against 11 general-purpose and energy-specific generation baselines. Experimental results show that SynEnergy improves anomaly preservation fidelity by an average of 12.21% and downstream quality by 2.96%, while maintaining competitive overall generation fidelity compared to baselines.

Figures

Figures reproduced from arXiv: 2608.03087 by the authors.

Figure 1
Figure 1. Spatial Distribution and Autocorrelation of Anom [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Similar Household Consumption Anomalies in Three Neighboring CBGs [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. The Overall Framework of SynEnergy. household sequences: e¯𝑛,𝑚 = e𝑛,𝑚 MAD(e𝑛,𝑚) + 𝜖𝑒 , (1) where MAD(⋅) denotes the median absolute deviation and 𝜖𝑒 en￾sures numerical stability. Since most residuals are near zero, we use soft weighting to emphasize informative deviations: w 𝑒 𝑛,𝑚 = sigmoid (𝛾𝑒 (⋃︀e¯𝑛,𝑚⋃︀ − 𝜏𝑒)) , e 𝑓 𝑛,𝑚 = w 𝑒 𝑛,𝑚 ⊙ e¯𝑛,𝑚, (2) where 𝜏𝑒 controls the activation position and 𝛾𝑒 controls the weight￾ing… view at source ↗
Figures from the paper (14 more)
Figure 5
Figure 5. Figure 5: The Layer-wise Control Injection Between Back [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Anomaly Rate [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 9
Figure 9. Figure 9: Q–Q Comparison of Anomalous Event Energy. 5.4 Anomalous Event Preservation (RQ5) To answer RQ5, we examine how well SynEnergy preserves two large-scale anomalous events in October 2018 and May 2019. In Fig￾ure 10, the red and blue lines represent the mean consumption o…
Figure 10
Figure 10. Figure 10: Visualization of the synthetic data generated by [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 11
Figure 11. Figure 11: Region Median In￾come [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 13
Figure 13. Figure 13: Region Higher Ed￾ucation Rate [PITH_FULL_IMAGE:figures/full_fig_p011_13.png]
Figure 15
Figure 15. Figure 15: Comparison of Zero-Consumption Rates [PITH_FULL_IMAGE:figures/full_fig_p012_15.png]
Figure 17
Figure 17. Figure 17: CCDF of Anomaly Event Magnitudes [PITH_FULL_IMAGE:figures/full_fig_p012_17.png]
Figure 20
Figure 20. Figure 20: Comparison of Sample-Level Anomaly-Rate Distributions. [PITH_FULL_IMAGE:figures/full_fig_p022_20.png]
Figure 21
Figure 21. Figure 21: Comparison of Sample-Level Anomalous Event Count Distributions. [PITH_FULL_IMAGE:figures/full_fig_p022_21.png]
Figure 22
Figure 22. Figure 22: Comparison of Sample-Level Anomalous Event Duration Distributions. [PITH_FULL_IMAGE:figures/full_fig_p022_22.png]
Figure 23
Figure 23. Figure 23: Comparison of Sample-Level Anomaly Magnitude Distributions. [PITH_FULL_IMAGE:figures/full_fig_p022_23.png]
Figure 24
Figure 24. Figure 24: Comparison of Sample-Level Anomaly Energy Distributions. [PITH_FULL_IMAGE:figures/full_fig_p023_24.png]
Figure 25
Figure 25. Figure 25: Violin-Plot Comparison of Sample-Level Anomaly Deviation Distributions. [PITH_FULL_IMAGE:figures/full_fig_p023_25.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 46 canonical work pages

  1. [1]

    Ahmed Alaa, Alex James Chan, and Mihaela van der Schaar. 2021. Generative time-series modeling with fourier flows. InInternational Conference on Learning Representations

    work page 2021

  2. [2]

    Michaël Allouche, Stéphane Girard, and Emmanuel Gobet. 2022. EV-GAN: Simu- lation of extreme events with ReLU neural networks.Journal of Machine Learning Research23, 150 (2022), 1–39

    work page 2022

  3. [3]

    Yihao Ang, Qiang Huang, Yifan Bao, Anthony KH Tung, and Zhiyong Huang. 2023. Tsgbench: Time series generation benchmark.arXiv preprint arXiv:2309.03755(2023)

    arXiv 2023

  4. [4]

    Luc Anselin. 2019. The Moran scatterplot as an ESDA tool to assess local insta- bility in spatial association. InSpatial analytical perspectives on GIS. Routledge, 111–126

    work page 2019

  5. [5]

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. 2016. Layer normaliza- tion.arXiv preprint arXiv:1607.06450(2016)

    arXiv 2016

  6. [6]

    Marin Biloš, Kashif Rasul, Anderson Schneider, Yuriy Nevmyvaka, and Stephan Günnemann. 2023. Modeling temporal data as continuous functions with sto- chastic process diffusion. InInternational Conference on Machine Learning. PMLR, 2452–2470

    work page 2023

  7. [7]

    Eoin Brophy, Zhengwei Wang, Qi She, and Tomás Ward. 2023. Generative adversarial networks in time series: A systematic literature review.Comput. Surveys55, 10 (2023), 1–31

    work page 2023

  8. [8]

    California Independent System Operator. 2024. Historical EMS Hourly Load. https://www.caiso.com/library/historical-ems-hourly-load

    work page 2024

Show all 58 references
  1. [9]

    Xueqi Cheng, Minxing Zheng, Shixiang Zhu, and Yushun Dong. 2025. Misleader: Defending against model extraction with ensembles of distilled models.arXiv preprint arXiv:2506.02362(2025)

    2025 arXiv

  2. [10]

    Abhyuday Desai, Cynthia Freeman, Zuhui Wang, and Ian Beaver. 2021. Timevae: A variational auto-encoder for multivariate time series generation.arXiv preprint arXiv:2111.08095(2021)

    2021 arXiv

  3. [11]

    Vivian Do, Heather McBrien, Nina M Flores, Alexander J Northrop, Jeffrey Schlegelmilch, Mathew V Kiang, and Joan A Casey. 2023. Spatiotemporal distri- bution of power outages with climate events and social vulnerability in the USA. Nature communications14, 1 (2023), 2470

    2023

  4. [12]

    2013.Modelling extremal events: for insurance and finance

    Paul Embrechts, Claudia Klüppelberg, and Thomas Mikosch. 2013.Modelling extremal events: for insurance and finance. Vol. 33. Springer Science & Business Media

    2013

  5. [13]

    Cristóbal Esteban, Stephanie L Hyland, and Gunnar Rätsch. 2017. Real-valued (medical) time series generation with recurrent conditional gans.arXiv preprint arXiv:1706.02633(2017)

    2017 arXiv

  6. [14]

    Marc Anton Finzi, Anudhyan Boral, Andrew Gordon Wilson, Fei Sha, and Leonardo Zepeda-Núñez. 2023. User-defined event sampling and uncertainty quantification in diffusion models for physical dynamical systems. InInternational Conference on Machine Learning. PMLR, 10136–10152

    2023

  7. [15]

    Chun Fu, Hussain Kazmi, Matias Quintana, and Clayton Miller. 2024. Creating synthetic energy meter data using conditional diffusion and building metadata. Energy and Buildings312 (2024), 114216

    2024

  8. [16]

    Michael Fuest, Alfredo Cuesta, and Kalyan Veeramachaneni. 2025. CENTS: Generating synthetic electricity consumption time series for rare and unseen scenarios.arXiv preprint arXiv:2501.14426(2025)

    2025 arXiv

  9. [17]

    Asadullah Hill Galib, Pang-Ning Tan, and Lifeng Luo. 2024. Fide: Frequency- inflated conditional diffusion model for extreme-aware time series generation. Advances in Neural Information Processing Systems37 (2024), 114434–114457

    2024

  10. [18]

    Ali Hasan, Khalil Elkhalil, Yuting Ng, João M Pereira, Sina Farsiu, Jose Blanchet, and Vahid Tarokh. 2022. Modeling extremes with 𝑑-max-decreasing neural networks. InUncertainty in Artificial Intelligence. PMLR, 759–768

    2022

  11. [19]

    Yi Hu, Yiyan Li, Lidong Song, Han Pyo Lee, PJ Rehm, Matthew Makdad, Edmond Miller, and Ning Lu. 2023. MultiLoad-GAN: A GAN-based synthetic load group generation method considering spatial-temporal correlations.IEEE Transactions on Smart Grid15, 2 (2023), 2309–2320

    2023

  12. [20]

    Lin Jiang, Dahai Yu, Ximiao Li, and Guang Wang. 2026. E4GEN: Event- level Explainable Extreme-Enhanced Time-series Generation.arXiv preprint arXiv:2606.01634(2026)

    2026 arXiv

  13. [21]

    Xuyuan Kang, Jingjing An, and Da Yan. 2023. A systematic review of building electricity use profile models.Energy and Buildings281 (2023), 112753

    2023

  14. [22]

    Marcel Kollovieh, Abdul Fatir Ansari, Michael Bohlke-Schneider, Jasper Zschieg- ner, Hao Wang, and Yuyang Bernie Wang. 2023. Predict, refine, synthesize: Self-guiding diffusion models for probabilistic time series forecasting.Advances in Neural Information Processing Systems36...

    2023

  15. [23]

    Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. 2021. DiffWave: A Versatile Diffusion Model for Audio Synthesis. InInternational Conference on Learning Representations

    2021

  16. [24]

    Lincan Li, Zheng Chen, and Yushun Dong. 2026. LLM as Clinical Graph Struc- ture Refiner: Enhancing Representation Learning in EEG Seizure Diagnosis. arXiv:2604.28178 [cs.AI] https://arxiv.org/abs/2604.28178

    2026 arXiv

  17. [25]

    Lincan Li, Zheng Chen, and Yushun Dong. 2026. NeuroGRIP: Retrieval- Augmented Graph Refinement for Knowledge-Grounded EEG Seizure Diagnosis. arXiv:2607.14314 [cs.LG] https://arxiv.org/abs/2607.14314

    2026 arXiv

  18. [26]

    Tianhong Li and Kaiming He. 2026. Back to basics: Let denoising generative models denoise. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 36115–36125

    2026

  19. [27]

    Yang Li, Han Meng, Zhenyu Bi, Ingolv T Urnes, and Haipeng Chen. 2025. Pop- ulation aware diffusion for time series generation. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 18520–18529

    2025

  20. [28]

    Ilan Naiman, Nimrod Berman, Itai Pemper, Idan Arbiv, Gal Fadlon, and Omri Azencot. 2024. Utilizing image transforms and diffusion models for generative modeling of short and long time series.Advances in Neural Information Processing Systems37 (2024), 121699–121730

    2024

  21. [29]

    Ilan Naiman, N Benjamin Erichson, Pu Ren, Michael W Mahoney, and Omri Azencot. 2024. Generative Modeling of Regular and Irregular Time Series Data via Koopman VAEs. InThe Twelfth International Conference on Learning Repre- sentations

    2024

  22. [30]

    Sai Shankar Narasimhan, Shubhankar Agarwal, Oguzhan Akcin, Sujay Sanghavi, and Sandeep Chinchali. 2024. Time weaver: A conditional time series generation model.arXiv preprint arXiv:2403.02682(2024)

    2024

  23. [31]

    New York Independent System Operator. 2024. Load Data. https://www.nyiso. com/load-data

    2024

  24. [32]

    Alexander Nikitin, Letizia Iannucci, and Samuel Kaski. 2024. TSGM: a flexible framework for generative modeling of synthetic time series.Advances in Neural Information Processing Systems37 (2024), 129042–129061

    2024

  25. [33]

    California Independent System Operator. 2024. Real-Time Load Data. https: //www.caiso.com/TodaysOutlook/Pages/default.aspx

    2024

  26. [34]

    New York Independent System Operator. 2024. Real-Time Load Data. https: //www.nyiso.com/load-data

    2024

  27. [35]

    Oreshkin, Dmitri Carpov, Nicolas Chapados, and Yoshua Bengio

    Boris N. Oreshkin, Dmitri Carpov, Nicolas Chapados, and Yoshua Bengio. 2020. N-BEATS: Neural Basis Expansion Analysis for Interpretable Time Series Fore- casting. InThe Eighth International Conference on Learning Representations (ICLR)

    2020

  28. [36]

    Kushagra Pandey, Jaideep Pathak, Yilun Xu, Stephan Mandt, Michael Pritchard, Arash Vahdat, and Morteza Mardani. 2025. Heavy-Tailed Diffusion Models. In International Conference on Learning Representations (ICLR)

    2025

  29. [37]

    William Peebles and Saining Xie. 2023. Scalable diffusion models with transform- ers. InProceedings of the IEEE/CVF international conference on computer vision. 4195–4205

    2023

  30. [38]

    Carlos Quesada, Leire Astigarraga, Chris Merveille, and Cruz E Borges. 2024. An electricity smart meter dataset of Spanish households: insights into consumption patterns.Scientific Data11, 1 (2024), 59

    2024

  31. [39]

    Mina Razghandi, Hao Zhou, Melike Erol-Kantarci, and Damla Turgut. 2023. Smart home energy management: VAE-GAN synthetic dataset generator and Q-learning. IEEE Transactions on Smart Grid15, 2 (2023), 1562–1573

    2023

  32. [40]

    Cécile Rousseau, Tobia Boschi, Giandomenico Cornacchia, Dhaval Salwala, Alessandra Pascale, and Juan Bernabe Moreno. 2025. Forging Time Series with Language: A Large Language Model Approach to Synthetic Data Generation. Advances in Neural Information Processing Systems(2025)

    2025

  33. [41]

    Sebastian Schmidl, Phillip Wenig, and Thorsten Papenbrock. 2022. Anomaly detection in time series: a comprehensive evaluation.Proceedings of the VLDB Endowment15, 9 (2022), 1779–1797

    2022

  34. [42]

    Vikash Sehwag, Caner Hazirbas, Albert Gordo, Firat Ozgenel, and Cristian Canton

  35. [43]

    Dario Shariatian, Umut Simsekli, and Alain Oliviero Durmus. 2025. Denoising levy probabilistic models. InInternational Conference on Learning Representations, Vol. 2025. 96991–97024

    2025

  36. [44]

    Bolin Shen, Md Shamim Seraj, Zhan Cheng, Shayok Chakraborty, and Yushun Dong. 2026. CITED: A Decision Boundary-Aware Signature for GNNs Towards Model Extraction Defense.arXiv preprint arXiv:2602.20418(2026)

    2026

  37. [45]

    Swapna Thorve, Young Yun Baek, Samarth Swarup, Henning Mortveit, Achla Marathe, Anil Vullikanti, and Madhav Marathe. 2023. High resolution synthetic residential energy use profiles for the United States.Scientific Data10, 1 (2023), 76

    2023

  38. [46]

    Census Bureau

    U.S. Census Bureau. 2020. American Community Survey 5-Year Data. https: //www.census.gov/programs-surveys/acs/data.html

    2020

  39. [47]

    U.S. EIA. 2020. Hourly electricity consumption varies throughout the day and across seasons. https://www.eia.gov/todayinenergy/detail.php?id=42915

    2020

  40. [48]

    Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research9, 11 (2008)

    2008

  41. [49]

    Chenxi Wang, Linxiao Yang, Zhixian Wang, Liang Sun, and Yi Wang. 2025. A Non-isotropic Time Series Diffusion Model with Moving Average Transitions. In Forty-second International Conference on Machine Learning

    2025

  42. [50]

    Martin B Wilk and Ram Gnanadesikan. 1968. Probability plotting methods for the analysis of data.Biometrika55, 1 (1968), 1–17

    1968

  43. [51]

    Tijin Yan, Hengheng Gong, He Yongping, Yufeng Zhan, and Yuanqing Xia. 2024. Probabilistic time series modeling with decomposable denoising diffusion model. InForty-first International Conference on Machine Learning

    2024

  44. [52]

    Jinsung Yoon, Daniel Jarrett, and Mihaela Van der Schaar. 2019. Time-series generative adversarial networks.Advances in neural information processing Conference’17, July 2017, Washington, DC, USA Lin Jiang, Dahai Yu, Ravikumar Gelli, and Guang Wang systems32 (2019)

    2019

  45. [53]

    Dahai Yu, Rongchao Xu, Lin Jiang, and Guang Wang. 2026. EnergyMamba: An Uncertainty-Aware Graph-Enhanced Selective State Space Model for Energy Consumption Prediction.arXiv preprint arXiv:2606.00506(2026)

    2026 arXiv

  46. [54]

    Dahai Yu, Rongchao Xu, Dingyi Zhuang, Yuheng Bu, Shenhao Wang, and Guang Wang. 2026. TrustEnergy: A Unified Framework for Accurate and Reliable User- level Energy Usage Prediction.Proceedings of the AAAI Conference on Artificial Intelligence40, 46 (Mar. 2026), 39558–39566. doi...

    2026 doi

  47. [55]

    Rui Yuan, S Ali Pourmousavi, Wen L Soong, Andrew J Black, Jon AR Liisberg, and Julian Lemos-Vinasco. 2023. A synthetic dataset of danish residential electricity prosumers.Scientific Data10, 1 (2023), 371

    2023

  48. [56]

    Xinyu Yuan and Yan Qiao. 2024. Diffusion-TS: Interpretable Diffusion for General Time Series Generation. InThe Twelfth International Conference on Learning Representations (ICLR)

    2024

  49. [57]

    Chad Zanocco, Tao Sun, Gregory Stelmach, June Flora, Ram Rajagopal, and Hilary Boudet. 2022. Assessing Californians’ awareness of their daily electricity use patterns.Nature Energy7, 12 (2022), 1191–1199. SynEnergy Conference’17, July 2017, Washington, DC, USA Figure 11: Regio...

    2022

  50. [2022]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Generating high fidelity data from low-density regions using diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 11492–11501

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.