WindFM: An Open-Source Foundation Model for Zero-Shot Wind Power Forecasting

Hang Fan; Jian Li; Shuo Chen; Wei Wei; Wei Xu; Yu Shi; Zongliang Fu

REVIEW 4 major objections 4 minor 40 references

WindFM: An Open-Source Foundation Model for Zero-Shot Wind Power Forecasting

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read WindFM claims a single 8.1M-parameter transformer, pre-trained on 150 billion simulated wind timesteps, can predict power at unseen wind farms with no fine-tuning, beating both site-trained models and much larger general-purpose foundation

desk verdict Domestic zero-shot claim is undermined by the site-count arithmetic, but the tokenizer and pretraining effort are worth a serious look. read the letter →

arxiv 2509.06311 v1 pith:V2VQ42LW submitted 2025-09-08 cs.LG

Hang Fan , Yu Shi , Zongliang Fu , Shuo Chen , Wei Wei , Wei Xu , Jian Li This is my paper · ORCID

classification cs.LG

keywords windpowerforecastingfoundationmodelzero-shotprobabilistictime-seriestokenizationbinarysphericalquantizationgenerativepre-trainingToolkit

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

WindFM's central claim is that wind power forecasting does not need a separately trained model for each wind farm: a single small model pre-trained on massive wind-energy data can forecast new sites directly. The authors build a discrete generative model in which continuous meteorological and power observations are converted into coarse and fine binary tokens, and a decoder-only Transformer pre-trains on roughly 150 billion time steps from more than 126,000 simulated U.S. sites. They report that the resulting 8.1M-parameter model, evaluated with no fine-tuning, beats both site-specific models trained on the target data and much larger general-purpose time-series foundation models on deterministic and probabilistic metrics. The paper also tests one out-of-distribution dataset from Inner Mongolia, China, and reports strong transfer. If the claim holds, wind farm operators could deploy a single open checkpoint to get calibrated point and probabilistic forecasts for new locations without per-site modeling.

What carries the argument

Discretize-and-generate framing with a coarse-to-fine token hierarchy. A BSQ tokenizer quantizes each continuous six-variate observation into a 20-bit binary token split into a 10-bit coarse subtoken and a 10-bit fine subtoken, replacing one prediction over a 2^20 vocabulary with two predictions over 2^10 vocabularies; a decoder-only Transformer then autoregressively predicts coarse-then-fine tokens, conditioned on Fourier-encoded timestamps. This machinery converts forecasting into token-sequence generation, makes probabilistic forecasts native through sampling, and is what lets an 8.1M-parameter model carry the reported generalization.

What would settle it

Run the released WindFM checkpoint zero-shot on a held-out real-world wind farm dataset from a region and climate unlike the U.S. simulation, for example an offshore North Sea site or a tropical low-wind site, and compare its MAE, RMSE, CRPS, and AQL against a site-specific model such as TimeMixer trained on that same site's own history. If WindFM's errors are not at least comparable to the site-trained model, the claim that it learns a universal, transferable representation of wind generation dynamics fails.

Watch

Extended reading notes

Core claim

The paper's discovery, stated on its own terms, is that a domain-specific discretize-and-generate pre-training pipeline yields a compact foundation model whose zero-shot forecasts transfer to unseen sites and unseen data frequencies. WindFM encodes the six-channel multivariate observation—wind power, wind speed, wind direction, air density, temperature, and pressure—into hierarchical discrete tokens via a Binary Spherical Quantization autoencoder: a coarse 10-bit subtoken captures dominant structure and a fine 10-bit subtoken captures residuals. A 4-layer decoder-only Transformer with Fourier temporal embeddings learns the transition distribution over these token sequences. After pre-trainin

Load-bearing premise

The load-bearing premise is that wind patterns in the WIND Toolkit, a simulated U.S. dataset, are representative enough of real wind farms worldwide that a model pre-trained on them can forecast an unseen site with no retraining; the paper tests this transfer on only one real out-of-distribution dataset.

Editorial extensions

If this is right

If the reported results hold, a wind farm operator can take the released checkpoint and produce zero-shot point forecasts for a new site without collecting a training set or tuning a per-site model.
The same inference pass can produce calibrated predictive intervals by drawing many samples, which is directly useful for reserve scheduling and electricity-market bidding.
Specialized pre-training on wind data would be a stronger recipe than scaling up a general time-series model, at least for this task, since WindFM's much smaller model reports lower errors.
Zero-shot performance on frequencies not seen in pre-training, such as 45-minute and 2-hour horizons, implies the learned dynamics are not tied to a specific sampling grid, a useful property for heterogeneous operational data.
Cross-continent transfer to the Inner Mongolia dataset, if representative, would make the model a plausible starting point for global wind forecasting without region-specific retraining.

Reading between the lines

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

Beyond the paper's tests, the same recipe could plausibly be applied to other energy variables such as solar photovoltaic output, load, and electricity price, since the tokenizer and autoregressive modeling do not use wind-specific physics; the authors list such fusion as future work, not a demonstrated result.
The reported dominance over site-specific models may depend on the tuned sampling recipe—temperature, top-p, and sample count—as much as on pre-training; a practical deployment would need to reproduce the inference-time settings, not just load the weights.
A direct probe of the learned tokens against physical regimes such as the power-curve plateau and cut-in or cut-out wind speeds would test whether the model's transfer comes from physics-like structure or from statistical co-occurrence; the paper does not include such an analysis.
The strongest extension would be a multi-continent out-of-distribution benchmark covering offshore, low-wind, or tropical sites; the single Inner Mongolia dataset is suggestive but too narrow to establish a universal representation globally.

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.

Desk Editor's Note

read the letter

Short version: the domestic zero-shot result doesn't survive contact with the paper's own site counts. 126,499 pre-training sites plus 192 validation sites equals 126,691, the entire WIND Toolkit. There are no sites left for the 7 'hold-out' sites in Section IV.B.1. So those 7 sites were almost certainly in the pre-training set, and the comparison is temporal generalization on seen sites, not zero-shot transfer to unseen locations. The only clean unseen-location test is Inner Mongolia, which is unreleased and not a clean win: WindFM loses RMSE at T=48 to TimeXer and at T=144 to TimeMOEbase. The headline 'unseen domestic sites' claim is therefore unsupported as written.

What's actually new and good: the hierarchical BSQ tokenizer with coarse and fine subtokens, and the chain-rule autoregressive prediction (first coarse, then fine conditioned on it), is a sensible combination that I don't see in the cited time-series foundation model literature. Pre-training on ~126k sites at multiple resolutions is a real effort, and the model is small (8.1M) and open-sourced. The experimental matrix is broad—site-specific and foundation baselines, four frequencies, three horizons each—and the sensitivity analysis of sampling hyperparameters is genuinely informative, showing the generative framing can be tuned for deterministic or probabilistic objectives.

Soft spots beyond the site count: inference hyperparameters (T, top-p, sample count) appear to be chosen from sensitivity curves computed on the test set, so the reported numbers are post-hoc selected. There are no error bars or repeated runs. The Inner Mongolia dataset isn't released, which matters because it's now the only clean OOD evidence. The ablations are useful but limited to one configuration each.

Who it's for: people working on time-series foundation models or wind forecasting will want to look at the tokenizer and the pretraining code, but they should not cite the zero-shot numbers until the site count and hyperparameter selection are clarified.

Recommendation: send it to peer review. The architecture and pretraining corpus are worth referee time, and the issues are fixable—release the OOD data, move hyperparameter tuning to a proper validation split, and correct the site arithmetic. A careful referee should be able to determine whether the domestic results change once held-out sites are actually held out.

Referee Report

4 major / 4 minor

Summary. The paper introduces WindFM, an 8.1M-parameter decoder-only transformer for zero-shot wind power forecasting. It discretizes multivariate meteorological-plus-power observations into hierarchical binary subtokens via Binary Spherical Quantization, then autoregressively pretrains on roughly 126,000 WIND Toolkit sites. Point forecasts are obtained as the mean of sampled trajectories; probabilistic forecasts are obtained from the empirical distribution of samples. The model is evaluated on four frequencies (15-min, 45-min, 1-hour, 2-hour), three horizons per frequency, against nine site-specific and seven foundation-model baselines, plus a cross-geography test on an Inner Mongolia dataset. The paper reports lower average MAE/RMSE/CRPS/AQL than baselines and emphasizes that WindFM achieves this without fine-tuning.

Significance. If the claims hold, WindFM is a practically valuable contribution: it demonstrates that a compact, domain-specific generative foundation model can outperform much larger general-purpose time-series foundation models and site-trained specialized models in zero-shot setting. The empirical scope is broad, with multiple frequencies, horizons, deterministic and probabilistic metrics, and an OOD geographical test. The release of the pretrained model is a concrete strength. However, the significance is conditional on resolving a site-count arithmetic inconsistency that affects the main domestic zero-shot claim, and on a clean protocol for choosing inference hyperparameters. The cross-geography evidence is also mixed on RMSE, so the absolute 'state-of-the-art' claim is stronger than what Table V supports.

major comments (4)

[Section III.B vs. Section IV.B.1] The site-count arithmetic contradicts the claim of '7 hold-out WIND Toolkit sites.' Section III.B says pretraining uses 126,499 sites and 192 sites are reserved for validation; Section II.B says the WIND Toolkit contains 126,691 sites. Since 126,499 + 192 = 126,691, no sites remain for the 7 hold-out sites evaluated in Section IV.B.1. If those 7 sites are part of the 126,499 pretraining sites, the model has seen their 2007-2012 data, so Tables I, II, and IV measure temporal generalization on seen sites, not zero-shot transfer to unseen locations. This is load-bearing for the paper's central contribution. Please provide site IDs and confirm they were excluded from pretraining, or revise the interpretation of the domestic zero-shot results.
[Section IV.A and Section IV.G (Figs. 7-8)] The zero-shot metrics reported in the main tables are obtained with inference hyperparameters (temperature, top-p, sample count) that are selected from sensitivity curves in Figs. 7-8. No separate validation set is described for this selection; if these curves are computed on the same test data used in Tables I-IV, the reported numbers are the result of test-set selection. The sensitivity is not negligible (MAE varies by roughly 10% across the swept range), and it can affect RMSE and probabilistic rankings. Please state explicitly which split the sensitivity analysis uses, or pre-specify the hyperparameters on a validation subset.
[Section IV.D, Table V] The abstract and Section IV.C claim state-of-the-art zero-shot performance on deterministic tasks. In the only clean unseen-location evaluation (Inner Mongolia), WindFM has the best MAE at all horizons but not the best RMSE: at T=48, TimeXer achieves RMSE 1.826 vs. WindFM 1.991; at T=144, TimeMOEbase achieves 1.715 vs. WindFM 1.744. This overstates the claim. Please either qualify the SOTA claim to MAE and aggregate domestic results, or report an average/across-horizon criterion that supports the statement.
[Section IV.B.3] The Inner Mongolia dataset is the only clean out-of-distribution test, but it is not described in enough detail (number of turbines, rated capacity, normalization of power values) and is not released. Since the domestic hold-out status is currently unclear, this dataset carries a large part of the generalization claim. The authors should make the dataset available or at least provide a full data card and a reproducible access path.

minor comments (4)

[Table II] The table formatting is broken in the submitted text (e.g., '1.0581.638', values running together). Please reformat for readability.
[Figure 6] The right panel shows downstream forecasting MAE/RMSE as a function of vocabulary size, but the experimental setting is not stated. Specify the dataset/split and whether this curve is on the test set; report standard errors or multiple runs.
[Section IV.B.1] For the site-specific baselines in the domestic zero-shot task, please state the exact train/validation/test split on the 7 sites and how the sites were chosen. This is needed to interpret the comparison.
[Section IV.D] Please report the number of turbines and the installed capacity or normalization factor for the Inner Mongolia dataset; otherwise the absolute MAE/RMSE values are not interpretable across datasets.

Circularity Check

2 steps flagged · score 6.0 of 10

Reported zero-shot metrics are partly selected on the test set via sampling-hyperparameter tuning, and the domestic 'hold-out site' split is contradicted by the paper's own site counts.

fitted input called prediction [Section IV.A (Inference hyperparameters) and Section IV.G (Sensitivity to Inference Hyperparameters)]
"Inference hyperparameters are tuned for each metric: for MAE, we use T= 0.6, top-p= 0.9, and 20 samples; for RMSE, which is more sensitive to large errors, we use a higher temperature of T= 0.9. For probabilistic forecasting, we generate an empirical distribution from 100 samples with T= 1.0 and top-p= 1.0 to maximize diversity."

The reported zero-shot metrics are not purely out-of-sample: temperature, top-p, and sample count are selected by sensitivity analysis on the evaluation data, and Figures 7-8 mark 'Optimal values used for evaluation.' The same metrics are then reported as WindFM's zero-shot performance. This is test-set fitting of inference constants: the headline MAE/RMSE/CRPS/AQL values are partly chosen to minimize the very errors they report, so the 'without any fine-tuning' claim obscures a small but real selection-on-targets step.
other [Section II.B, Section III.B.3, Section IV.B.1]
"The dataset contains meteorological and turbine power data from 126,691 sites across the continental United States ... for the years 2007–2013. ... The pre-training dataset is constructed from the WIND Toolkit, containing time series from 126,499 sites for the years 2007–2012. An additional 192 sites are reserved for validation. ... This task assesses zero-shot accuracy on unseen data ... (7 hold-out WIND Toolkit sites, year 2013)."

126,499 + 192 = 126,691, exactly the stated total number of WIND Toolkit sites, leaving no room for 7 additional hold-out sites. As written, the 7 domestic 'hold-out' sites must either lie inside the 126,499 pre-training sites or inside the 192 validation sites. If the former, the domestic zero-shot evaluation is temporal extrapolation on sites already seen in pre-training, not transfer to unseen locations; if the latter, the evaluation sites were used for validation/model selection. Either way, the central 'unseen domestic sites' claim is not supported as stated, and the only clean unseen-location evidence is the single unreleased Inner Mongolia dataset.

full rationale

The core tokenizer and autoregressive training derivation are self-contained: the discrete hierarchical tokenization, the chain-rule factorization in Eqs. (3)-(4), and the negative log-likelihood objective in Eq. (9) do not reduce to the evaluation targets, and the method is benchmarked against external baselines and an OOD dataset. The self-citation to Kronos [1] is motivational only and not load-bearing. However, two issues create partial circularity in the reported zero-shot claims. First, the inference sampling hyperparameters (temperature, top-p, sample count) are explicitly 'tuned for each metric' and the sensitivity analysis marks 'Optimal values used for evaluation,' meaning the reported zero-shot numbers are partially selected on the test set. Second, the paper's own site counts (126,499 pre-training + 192 validation = 126,691 total) leave no room for the 7 'hold-out WIND Toolkit sites,' so the domestic unseen-site evaluation is either not unseen or relies on validation-set leakage. These problems affect the headline zero-shot SOTA claim, though the model's architecture and OOD Inner Mongolia results retain independent content. Score 6 reflects partial circularity: one reported prediction is partly fitted via test-set-tuned constants, and the other is undermined by the data-split arithmetic.

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

The central claim is an empirical performance claim; the free parameters are mostly sampling hyperparameters tuned on the evaluation set, which is the main circularity burden. The axioms are domain assumptions about the representativeness of the training data and the sufficiency of the discrete representation. No invented entities were introduced.

free parameters (5)

Sampling temperature T = 0.6 (deterministic MAE), 0.9 (deterministic RMSE), 1.0 (probabilistic)
Tuned on the test set via sensitivity analysis (Figs. 7-8); this is a direct fit to the evaluation data, inflating zero-shot scores.
Top-p (nucleus sampling) = 0.9 (deterministic), 1.0 (probabilistic)
Chosen based on test-set sensitivity curves (Fig. 7b, 8b).
Sample count = 20 (deterministic), 100 (probabilistic)
Selected from diminishing-return curves on the test set (Fig. 7c, 8c).
Tokenizer bit length k = k=20 (k1=k2=10)
Selected from reconstruction and forecasting trade-off in Fig. 6; hand-chosen architecture hyperparameter.
Tokenizer loss weights = beta=0.05, gamma0=1.0, gamma=1.1, zeta=0.05, lambda=1.0
Hand-set values inherited from BSQ/LFQ implementations, not tuned per task.

assumptions (4)

domain assumption The NREL WIND Toolkit is representative of real-world wind dynamics.
Pre-training uses only simulated US data; the assumption that the learned dynamics transfer to real operational farms is tested on just one Chinese dataset (Sec. IV-D).
ad hoc to paper Discrete tokenization with k=20 bits preserves forecasting-relevant information.
Bits chosen empirically via Fig. 6; no guarantee that discretization error does not dominate forecast error for extreme events.
domain assumption A decoder-only transformer can learn a universal representation of wind dynamics across sites and resolutions.
Standard assumption for foundation models; empirically validated on hold-out sites.
ad hoc to paper Autoregressive generation of the full multivariate vector (including covariates) is a valid forecasting strategy.
The model must generate future wind speed, direction, etc. without external NWP inputs, an unusual design choice.

how reviews work

0 comments

Cite this review

Pith. "Pith review of WindFM: An Open-Source Foundation Model for Zero-Shot Wind Power Forecasting." pith.science (2026). https://pith.science/paper/V2VQ42LW

@misc{pith2026250906311,
  author       = {Pith},
  title        = {Pith review of: WindFM: An Open-Source Foundation Model for Zero-Shot Wind Power Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V2VQ42LW}},
  note         = {Machine review of arXiv:2509.06311}
}

read the original abstract

High-quality wind power forecasting is crucial for the operation of modern power grids. However, prevailing data-driven paradigms either train a site-specific model which cannot generalize to other locations or rely on fine-tuning of general-purpose time series foundation models which are difficult to incorporate domain-specific data in the energy sector. This paper introduces WindFM, a lightweight and generative Foundation Model designed specifically for probabilistic wind power forecasting. WindFM employs a discretize-and-generate framework. A specialized time-series tokenizer first converts continuous multivariate observations into discrete, hierarchical tokens. Subsequently, a decoder-only Transformer learns a universal representation of wind generation dynamics by autoregressively pre-training on these token sequences. Using the comprehensive WIND Toolkit dataset comprising approximately 150 billion time steps from more than 126,000 sites, WindFM develops a foundational understanding of the complex interplay between atmospheric conditions and power output. Extensive experiments demonstrate that our compact 8.1M parameter model achieves state-of-the-art zero-shot performance on both deterministic and probabilistic tasks, outperforming specialized models and larger foundation models without any fine-tuning. In particular, WindFM exhibits strong adaptiveness under out-of-distribution data from a different continent, demonstrating the robustness and transferability of its learned representations. Our pre-trained model is publicly available at https://github.com/shiyu-coder/WindFM.

Figures

Figures reproduced from arXiv: 2509.06311 by the authors.

**Figure 2.** Visualization of complex wind dynamics from a representative site. (Left) Six 3D plots reveal strong diurnal and seasonal cycles in meteorological [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** The overall framework of WindFM, consisting of two main stages: (1) Time-series Tokenization, which converts continuous multivariate series into [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: Architecture of the Time-series Tokenizer. It utilizes a Transformer [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Zero-shot forecasting performance versus model size. WindFM (red) [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Impact of tokenizer vocabulary size (2 k) on performance. Increasing the vocabulary size improves the model’s representational capacity, leading to a consistent decrease in both time-series reconstruction error (left) and downstream power forecasting error (right) in t…

**Figure 7.** Figure 7: Sensitivity analysis of WindFM’s forecasting performance with respect to inference sampling hyperparameters. We evaluate MAE and RMSE by: (a) [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

**Figure 8.** Figure 8: Sensitivity analysis of WindFM’s probabilistic forecasting performance with respect to inference sampling hyperparameters. We evaluate CRPS and [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

**Figure 9.** Figure 9: Visualization of WindFM’s zero-shot forecasts against other baselines. The plots cover three different time frequencies for an unseen wind farm. [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]

**Figure 10.** Figure 10: Visualization of WindFM’s zero-shot probabilistic forecasts. The plots display the median forecast, alongside the 50% and 90% confidence intervals, [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 26 canonical work pages

[1]

Kronos: A foundation model for the language of financial markets,

Y . Shi, Z. Fu, S. Chenet al., “Kronos: A foundation model for the language of financial markets,”arXiv preprint arXiv:2508.02739, 2025

arXiv 2025
[2]

Image and video tokenization with binary spherical quantization,

Y . Zhao, Y . Xiong, and P. Kr ¨ahenb¨uhl, “Image and video tokenization with binary spherical quantization,”arXiv preprint arXiv:2406.07548, 2024

arXiv 2024
[3]

Language model beats diffusion–tokenizer is key to visual generation,

L. Yu, J. Lezama, N. B. Gundavarapuet al., “Language model beats diffusion–tokenizer is key to visual generation,”arXiv preprint arXiv:2310.05737, 2023

arXiv 2023
[4]

Review of meta-heuristic algorithms for wind power prediction: Methodologies, applications and challenges,

P. Lu, L. Ye, Y . Zhaoet al., “Review of meta-heuristic algorithms for wind power prediction: Methodologies, applications and challenges,” Applied Energy, vol. 301, p. 117446, 2021

work page 2021
[5]

The wind integration national dataset (wind) toolkit,

C. Draxl, A. Clifton, B.-M. Hodgeet al., “The wind integration national dataset (wind) toolkit,”Applied Energy, vol. 151, pp. 355–366, 2015

work page 2015
[6]

Global wind report 2024,

Global Wind Energy Council, “Global wind report 2024,” GWEC, Tech. Rep., 2024. [Online]. Available: https://gwec.net/ global-wind-report-2024/

work page 2024
[7]

WWEA annual report 2024: A challenging year for windpower,

World Wind Energy Association, “WWEA annual report 2024: A challenging year for windpower,” WWEA, Tech. Rep., 2024. [Online]. Available: https://www.wwindea.org/AnnualReport2024

work page 2024
[8]

A survey on wind power forecasting with machine learning approaches,

Y . Yang, H. Lou, J. Wuet al., “A survey on wind power forecasting with machine learning approaches,”Neural Computing and Applications, vol. 36, no. 21, pp. 12 753–12 773, 2024

work page 2024

Show all 40 references

[9]

Deep learning model-transformer based wind power forecasting approach,

S. Huang, C. Yan, and Y . Qu, “Deep learning model-transformer based wind power forecasting approach,”Front. Energy Res., vol. 10, p. 1055683, 2023

2023
[10]

Chronos: Learning the language of time series,

A. F. Ansari, L. Stella, C. Turkmenet al., “Chronos: Learning the language of time series,”arXiv preprint arXiv:2403.07815, 2024

2024 arXiv
[11]

Time-moe: Billion-scale time series foundation models with mixture of experts,

X. Shi, S. Wang, Y . Nieet al., “Time-moe: Billion-scale time series foundation models with mixture of experts,”arXiv preprint arXiv:2409.16040, 2024

2024 arXiv
[12]

Temporal collaborative attention for wind power forecasting,

Y . Hu, H. Liu, S. Wuet al., “Temporal collaborative attention for wind power forecasting,”Applied Energy, vol. 357, p. 122502, 2024

2024
[13]

itransformer: Inverted transformers are effective for time series forecasting,

Y . Liu, T. Hu, H. Zhanget al., “itransformer: Inverted transformers are effective for time series forecasting,”arXiv preprint arXiv:2310.06625, 2023

2023 arXiv
[14]

Fedformer: Frequency enhanced de- composed transformer for long-term series forecasting,

T. Zhou, Z. Ma, Q. Wenet al., “Fedformer: Frequency enhanced de- composed transformer for long-term series forecasting,” inInternational Conference on Machine Learning. PMLR, 2022, pp. 27 268–27 286

2022
[15]

Timexer: Empowering transformers for time series forecasting with exogenous variables,

Y . Wang, H. Wu, J. Donget al., “Timexer: Empowering transformers for time series forecasting with exogenous variables,”Advances in Neural Information Processing Systems, vol. 37, pp. 469–498, 2024

2024
[16]

A time series is worth 64 words: Long-term forecasting with transformers,

Y . Nie, N. H. Nguyen, P. Sinthonget al., “A time series is worth 64 words: Long-term forecasting with transformers,”arXiv preprint arXiv:2211.14730, 2022

2022 arXiv
[17]

Non-stationary transformers: Exploring the stationarity in time series forecasting,

Y . Liu, H. Wu, J. Wanget al., “Non-stationary transformers: Exploring the stationarity in time series forecasting,”Advances in neural informa- tion processing systems, vol. 35, pp. 9881–9893, 2022

2022
[18]

Timemixer: Decomposable multiscale mixing for time series forecasting,

S. Wang, H. Wu, X. Shiet al., “Timemixer: Decomposable multiscale mixing for time series forecasting,”arXiv preprint arXiv:2405.14616, 2024

2024 arXiv
[19]

Timesnet: Temporal 2d-variation modeling for general time series analysis,

H. Wu, T. Hu, Y . Liuet al., “Timesnet: Temporal 2d-variation modeling for general time series analysis,”arXiv preprint arXiv:2210.02186, 2022

2022 arXiv
[20]

Are transformers effective for time series forecasting?

A. Zeng, M. Chen, L. Zhanget al., “Are transformers effective for time series forecasting?” inProceedings of the AAAI Conference on Artificial Intelligence, no. 9, 2023, pp. 11 121–11 128

2023
[21]

Moment: A family of open time-series foundation models,

M. Goswami, K. Szafer, A. Choudhryet al., “Moment: A family of open time-series foundation models,”arXiv preprint arXiv:2402.03885, 2024

2024 arXiv
[22]

A decoder-only foundation model for time-series forecasting,

A. Das, W. Kong, R. Senet al., “A decoder-only foundation model for time-series forecasting,” inForty-first International Conference on Machine Learning, 2024

2024
[23]

Unified training of universal time series forecasting transformers,

G. Woo, C. Liu, A. Kumaret al., “Unified training of universal time series forecasting transformers,”arXiv preprint arXiv:2402.02592, 2024

2024 arXiv
[24]

Roformer: Enhanced transformer with rotary position embedding,

J. Su, M. Ahmed, Y . Luet al., “Roformer: Enhanced transformer with rotary position embedding,”Neurocomputing, vol. 568, p. 127063, 2024

2024
[25]

On layer normalization in the trans- former architecture,

R. Xiong, Y . Yang, D. Heet al., “On layer normalization in the trans- former architecture,” inInternational conference on machine learning. PMLR, 2020, pp. 10 524–10 533

2020
[26]

Root mean square layer normalization,

B. Zhang and R. Sennrich, “Root mean square layer normalization,” Adv. Neural Inf. Process. Syst., vol. 32, 2019

2019
[27]

Decoupled weight decay regularization,

I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017

2017 arXiv
[28]

Empower pre-trained large language models for building-level load forecasting,

Y . Zhou and M. Wang, “Empower pre-trained large language models for building-level load forecasting,”IEEE Transactions on Power Systems, 2025

2025
[29]

Exploring the capabilities and limitations of large language models in the electric energy sector,

S. Majumder, L. Dong, F. Doudiet al., “Exploring the capabilities and limitations of large language models in the electric energy sector,”Joule, vol. 8, no. 6, pp. 1544–1549, 2024

2024
[30]

High and low frequency wind power prediction based on transformer and bigru-attention,

S. Wang, J. Shi, W. Yanget al., “High and low frequency wind power prediction based on transformer and bigru-attention,”Energy, vol. 288, p. 129753, 2024

2024
[31]

Multi-source and temporal attention network for probabilistic wind power prediction,

H. Zhang, J. Yan, Y . Liuet al., “Multi-source and temporal attention network for probabilistic wind power prediction,”IEEE Transactions on Sustainable Energy, vol. 12, no. 4, pp. 2205–2218, 2021

2021
[32]

Augmented convolutional network for wind power prediction: A new recurrent architecture design with spatial- temporal image inputs,

L. Cheng, H. Zang, Y . Xuet al., “Augmented convolutional network for wind power prediction: A new recurrent architecture design with spatial- temporal image inputs,”IEEE Transactions on Industrial Informatics, vol. 17, no. 10, pp. 6981–6993, 2021

2021
[33]

Eplus-llm: A large language model- based computing platform for automated building energy modeling,

G. Jiang, Z. Ma, L. Zhanget al., “Eplus-llm: A large language model- based computing platform for automated building energy modeling,” Appl. Energy, vol. 367, p. 123431, 2024

2024
[34]

Continuous and distribution-free probabilistic wind power forecasting: A conditional normalizing flow approach,

H. Wen, P. Pinson, J. Maet al., “Continuous and distribution-free probabilistic wind power forecasting: A conditional normalizing flow approach,”IEEE Transactions on Sustainable Energy, vol. 13, no. 4, pp. 2250–2263, 2022

2022
[35]

A novel frequency-domain physics-informed neural network for accurate prediction of 3d spatio-temporal wind fields in wind turbine applications,

S. Li, X. Li, Y . Jianget al., “A novel frequency-domain physics-informed neural network for accurate prediction of 3d spatio-temporal wind fields in wind turbine applications,”Applied Energy, vol. 386, p. 125526, 2025

2025
[36]

Bert4st:: Fine-tuning pre-trained large language model for wind power forecasting,

Z. Lai, T. Wu, X. Feiet al., “Bert4st:: Fine-tuning pre-trained large language model for wind power forecasting,”Energy Conv. Manag., vol. 307, p. 118331, 2024

2024
[37]

Pricefm: Foundation model for prob- abilistic electricity price forecasting,

R. Yu, C. Gu, J. Stiasnyet al., “Pricefm: Foundation model for prob- abilistic electricity price forecasting,”arXiv preprint arXiv:2508.04875, 2025

2025 arXiv
[38]

A novel genetic lstm model for wind power forecast,

F. Shahid, A. Zameer, and M. Muneeb, “A novel genetic lstm model for wind power forecast,”Energy, vol. 223, p. 120069, 2021

2021
[39]

Wind power fore- casting enhancement utilizing adaptive quantile function and cnn-lstm: a probabilistic approach,

O. Abedinia, A. Ghasemi-Marzbali, M. Shafieiet al., “Wind power fore- casting enhancement utilizing adaptive quantile function and cnn-lstm: a probabilistic approach,”IEEE Transactions on Industry Applications, vol. 60, no. 3, pp. 4446–4457, 2024

2024
[40]

A novel photovoltaic power probabilistic forecasting model based on monotonic quantile convolutional neural network and multi-objective optimization,

J. Zhu and Y . He, “A novel photovoltaic power probabilistic forecasting model based on monotonic quantile convolutional neural network and multi-objective optimization,”Energy Conversion and Management, vol. 323, p. 119219, 2025

2025

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Kronos: A foundation model for the language of financial markets,

Y . Shi, Z. Fu, S. Chenet al., “Kronos: A foundation model for the language of financial markets,”arXiv preprint arXiv:2508.02739, 2025

arXiv 2025

[2] [2]

Image and video tokenization with binary spherical quantization,

Y . Zhao, Y . Xiong, and P. Kr ¨ahenb¨uhl, “Image and video tokenization with binary spherical quantization,”arXiv preprint arXiv:2406.07548, 2024

arXiv 2024

[3] [3]

Language model beats diffusion–tokenizer is key to visual generation,

L. Yu, J. Lezama, N. B. Gundavarapuet al., “Language model beats diffusion–tokenizer is key to visual generation,”arXiv preprint arXiv:2310.05737, 2023

arXiv 2023

[4] [4]

Review of meta-heuristic algorithms for wind power prediction: Methodologies, applications and challenges,

P. Lu, L. Ye, Y . Zhaoet al., “Review of meta-heuristic algorithms for wind power prediction: Methodologies, applications and challenges,” Applied Energy, vol. 301, p. 117446, 2021

work page 2021

[5] [5]

The wind integration national dataset (wind) toolkit,

C. Draxl, A. Clifton, B.-M. Hodgeet al., “The wind integration national dataset (wind) toolkit,”Applied Energy, vol. 151, pp. 355–366, 2015

work page 2015

[6] [6]

Global wind report 2024,

Global Wind Energy Council, “Global wind report 2024,” GWEC, Tech. Rep., 2024. [Online]. Available: https://gwec.net/ global-wind-report-2024/

work page 2024

[7] [7]

WWEA annual report 2024: A challenging year for windpower,

World Wind Energy Association, “WWEA annual report 2024: A challenging year for windpower,” WWEA, Tech. Rep., 2024. [Online]. Available: https://www.wwindea.org/AnnualReport2024

work page 2024

[8] [8]

A survey on wind power forecasting with machine learning approaches,

Y . Yang, H. Lou, J. Wuet al., “A survey on wind power forecasting with machine learning approaches,”Neural Computing and Applications, vol. 36, no. 21, pp. 12 753–12 773, 2024

work page 2024