Pith. sign in

REVIEW 3 major objections 6 minor 60 references

Pairing turbine logs with gridded weather forecasts and fusing them in the frequency domain yields the most accurate short-term wind power forecasts on three real farms.

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 · deepseek-v4-flash

2026-08-01 19:02 UTC pith:ZAL2B3MT

load-bearing objection A competent, genuinely new fusion architecture for wind power forecasting, but the CERRA 'forecast' provenance is unresolved and the hyperparameter selection may be peeking at the test set. the 3 major comments →

arxiv 2607.17095 v1 pith:ZAL2B3MT submitted 2026-07-19 cs.AI

Fourier Geometric Wind Power Forecasting with Numerical Weather Prediction

Shiyuan Piao , Fan Zehui , Yang Liu , Hong Cheng , Juepeng Zheng , Jie Zhou , Fugee Tsung This is my paper
classification cs.AI
keywords wind power forecastingnumerical weather predictionFourier neural operatorSCADA datamultimodal fusiongeometric deep learningrotation invarianceshort-term forecasting
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.

Short-term wind power forecasting is hard because turbine output depends on both the turbine's recent state and atmospheric conditions that have not yet arrived. The paper tries to establish that the right way to combine these two information sources is a multimodal model that (1) turns raw wind vectors and nacelle angles into rotation-invariant features—yaw misalignment, horizontal divergence, vorticity, mean advection—and (2) fuses point-level turbine history with gridded weather forecasts using learnable spectral mixing in the Fourier domain along the time axis. On three UK wind farms with forecasts up to six hours ahead, the model reports the lowest MAE and RMSE among a wide set of baselines, with roughly 19–32% lower errors than the strongest SCADA-only models and 15–29% lower than SCADA-plus-weather fusion baselines. Ablations attribute the largest degradations to removing the weather input and removing the spectral fusion layer. If the claim holds, it gives grid operators a concrete, parameter-efficient way to use forecast weather fields for dispatch and reserve decisions at the horizon where NWP still has skill.

Core claim

The central claim is that the bottleneck in short-term wind power forecasting is not the availability of weather forecasts but how they are fused with turbine data. The framework replaces raw directional inputs with deterministic geometric features—wind-to-nacelle misalignment angle, power ramp rate, domain-mean wind vector, divergence, vorticity—then projects turbine and grid embeddings into a shared space and processes them with Fourier Neural Operator layers. These layers apply a discrete Fourier transform along time, mix real and imaginary spectral parts with learnable weights, and transform back before concatenating turbine and grid nodes; the decoder is a two-layer MLP. Experiments cla

What carries the argument

The load-bearing mechanism is the frequency-domain fusion module: it takes each modality's embedding, applies a discrete Fourier transform along the time axis, separates the spectrum into real and imaginary parts, applies a learnable linear mixing in the frequency domain, and returns to the time domain via inverse transform. Because the spectral weights are shared across turbines and grid points, this performs a global temporal convolution that can align slowly evolving weather fields with fast turbine-level power fluctuations. The second mechanism is the non-parametric geometric encoder, which enforces rotation invariance before any learning: yaw misalignment is computed as the angle betwee

Load-bearing premise

The load-bearing premise is that the weather data labelled as forecasts for the next one-to-six hours really are forecasts made at the starting time, not records that already include what actually happened afterward. The paper's data description mixes these two kinds of weather fields and does not name the exact product used for the forecast part.

What would settle it

Take any held-out test sample and compare the model's forecast-input grid with the weather forecast that was actually issued at that hour and with the analysis for the same hour. If the input tracks the analysis more closely than the genuine forecast, or if replacing forecast lead times with lagged pre-forecast analyses preserves the reported accuracy gain, the central claim that future weather information drives the improvement is falsified.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the framework's claims hold, operational forecasters can expect the largest accuracy gains at the 3–6 hour horizon, exactly where SCADA-only extrapolation degrades and where NWP still has useful lead-time signal.
  • The ablations imply that future weather information and spectral fusion are both necessary: removing either roughly doubles the error increase seen from removing geometric priors, so a model with only one of the two would be substantially weaker.
  • The reported error reductions translate to roughly 5.8–7.2 kW per turbine-step per 0.01 normalized MAE, so gains accumulate over turbines and rolling dispatch windows into reduced imbalance risk.
  • The architecture is lightweight (about 0.12M parameters in the decoder) and resolution-flexible, suggesting it can be applied to other farms without heavy per-site tuning.
  • Cross-site experiments on two additional farms show smaller but consistent gains, indicating the design transfers beyond the three main sites.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's ablations show that removing the weather input hurts, but they do not isolate whether the benefit comes from future information or from having any gridded weather covariates; a control using only pre-forecast analysis fields would settle which mechanism the paper's story depends on.
  • Because the geometric encoder only cares about relative angles, the same design could be applied to other direction-sensitive renewable forecasting problems, such as solar panels with orientation-dependent irradiance; the paper does not claim this.
  • The frequency-domain fusion along time should be indifferent to the number of grid points, so the architecture could in principle handle higher-resolution NWP grids or longer horizons without restructuring; the paper only tests fixed 9-step history and 6-step horizon.
  • The reported kilowatt-per-turbine conversions suggest operational value, but the paper stops short of pricing errors under real imbalance tariffs; a cost-weighted evaluation would show whether the remaining MAE gaps are commercially decisive.

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 GWPF, a multimodal wind power forecasting framework that combines turbine-level SCADA history with gridded CERRA NWP fields for 1–6 h ahead prediction. Inputs are decomposed into scalar and vector features; a non-parametric geometric encoder computes rotation-invariant quantities such as yaw misalignment, divergence, vorticity, and mean advection; and a temporal Fourier-style spectral fusion module mixes the two modalities. Experiments on three UK wind farms report consistent MAE/RMSE improvements over SCADA-only, SCADA+CERRA fusion, and NWP-only baselines, with ablations attributing the gains to the CERRA input, the learnable spectral fusion, and the geometric priors. The code is publicly available.

Significance. If the results are reproducible under genuinely prospective NWP inputs, the paper makes a useful contribution: a physics-inspired geometric encoder for wind vectors and yaw misalignment, a parameter-efficient temporal spectral fusion mechanism, a public implementation, and a controlled comparison against fusion baselines that share the same inputs. The ablations are informative and mostly support the design. The main risks are the unverified temporal validity of the CERRA 'forecast' fields and the per-dataset selection of the number of fusion layers; both bear directly on the validity of the headline numbers.

major comments (3)
  1. [§5.1 and §9.1] The paper does not specify which CERRA product supplies the 'short-range forecasts' used for lead times 1–6 h. §5.1 says the dataset includes 'both reanalysis (analysis) and short-range forecasts,' §9.1 says 'forecasts issued at t0 for lead times 1–6 hours,' but §7 refers to all data as 'CERRA reanalysis data.' The cited dataset [41] is a regional reanalysis archive, not an operational forecast archive. If the future fields are analysis fields valid at the lead time, or were produced with later observations in the assimilation, then future target information leaks into the model inputs at train and test time. In that case the claimed 19–32% gains versus SCADA-only baselines and 15–29% versus SCADA+CERRA fusion baselines are not operationally achievable, and the w/o CERRA ablation cannot separate 'future weather helps' from 'target information leaked.' Please identify the exact CERRA prod
  2. [Table 6 and §9.2] The number of NFL layers is selected per dataset in Table 6, but no validation-set criterion is stated. The final 'Ours' results in Tables 1–4 appear to use the per-dataset optimum (NFL=1 for Kelmarsh, 2 for Penmanshiel, 3 for Hill of Towie), while §9.2 says the final model uses 'two frequency-fusion layers.' If the test set was used to choose the depth, the reported results are optimistically biased and the comparison to baselines (which likely use a fixed configuration) is unfair. Specify the validation split used for model selection, report the selected depths, and reconcile the discrepancy with §9.2. Ideally, keep the architecture depth fixed across datasets or show that results are stable across a small range.
  3. [§5.2 / Table 2 caption] The fusion baseline protocol is described inconsistently. The main text (§5.3) says Table 2 baselines use a shared Informer-style token-concatenation embedding, while §9.3 says 'Other SCADA+CERRA fusion baselines' use a shared GRU for SCADA histories plus an MLP for CERRA forecasts. This discrepancy makes it difficult to reconstruct the controlled comparison and to determine whether the reported gap over fusion baselines is due to the proposed architecture or to suboptimal baseline input encoding. Please align the descriptions and provide the exact embedding details used for Table 2.
minor comments (6)
  1. [§5.2] The sentence 'All baselines in table 2 are only input the scada data' should refer to Table 1; as written it contradicts the table's caption and the subsequent discussion. Please correct.
  2. [Table 8] The note 'Input/Pred. = 3h history/1h horizon' conflicts with the stated prediction horizon H=6 and with Table 8's 'Pred. 6' column. Correct to '6h horizon'.
  3. [Table 5] The features 'wake tendency' and 'turbulence intensity' are used in the fine-grained analysis but never defined. Provide formulas or references.
  4. [§5.1, target mask] The validity mask uses the future 'Lost Production' flag; in an operational setting this information may not be available at t0. This should be stated as a limitation when interpreting the absolute error levels (e.g., in §5.3's operational conversion).
  5. [§4.3 and Abstract] The paper claims FNO captures 'long-range spatiotemporal relationships,' but the frequency fusion is applied along the time axis only; spatial mixing occurs in the final MLP decoder. Consider rephrasing to avoid overclaiming spatial spectral convolution.
  6. [Abstract] Typo: 'We further leverages' should be 'We further leverage.'

Circularity Check

0 steps flagged

No circular derivation: performance claims rest on held-out empirical tests and deterministic physical feature transforms; the CERRA forecast-provenance ambiguity is a correctness risk, not a circularity.

full rationale

The paper is an empirical ML study: Eq. (9) trains the model by masked MSE against held-out 2020 turbine power, and the headline claims are test-set MAE/RMSE numbers. I find no step in which a 'prediction' is equivalent by construction to a fitted input. The geometric encoder (Eqs. 10–16) is a non-parametric, deterministic map from wind vectors, yaw, and coordinates to alignment/divergence/vorticity/advection features; these features are not fit to the target and therefore cannot smuggle the answer in. The FNO-based fusion (Eq. 17 and the M module) is a standard spectral mixing layer with learnable weights trained under the same loss; selecting the number of NFL layers in Table 6 is ordinary hyperparameter/model selection, not fitting a parameter to the predicted quantity. The authors' self-citations ([29,31,32,56,57]) appear in the related-work overview as examples of GNN/equivariant forecasting and are not used as the load-bearing justification for this paper's design, nor is any uniqueness theorem imported. The only serious concern—that the CERRA 'short-range forecasts' at lead times 1–6 h might not be genuine t0 forecasts but analysis fields incorporating later observations—would undermine the forecasting validity and inflate the reported gains. That is a data-integrity/correctness risk raised by the text's own inconsistency (§7 calls all data 'CERRA reanalysis data' while §5.1/§9.1 distinguish analysis history from t0 forecasts), but it is not a circular derivation: no equation or fitted parameter reduces the headline result to the inputs by construction. Because the central claim has independent empirical content and no circular step is exhibitable, the circularity score is low (2), reflecting the non-load-bearing self-citations and the unresolved data-provenance caveat.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central claim rests on no invented physical entities. It relies on standard Fourier mathematics, domain physics encoded as hand-built features, and assumptions about data provenance and the representativeness of the masking rule. The main free choices are architecture hyperparameters and the per-farm NFL depth, with the latter being the most consequential because it may have been selected on the test set.

free parameters (3)
  • Number of NFL fusion layers = 1 (Kelmarsh), 2 (Penmanshiel), 3 (Hill of Towie)
    Selected per dataset as the best test MAE in Table 6; the paper does not state that this selection was made on validation data, so the reported main results may incorporate test-set information.
  • Embedding dimensions d_time, d_hidden = 16, 8
    Hand-chosen in Appendix 9.2; not swept or justified, and the model's capacity depends on them.
  • Target validity mask (Lost Production == 0 and finite power) = binary exclusion rule
    A hand-chosen data-exclusion rule used in every metric and loss; if curtailment/downtime correlates with hard-to-forecast regimes, it could inflate apparent accuracy.
axioms (5)
  • standard math The Discrete Fourier Transform pair (Eqs. 2–4) and the FNO layer update (Eq. 5) are valid for the discrete inputs considered.
    Used in §§3.2 and 4.3; no discussion of aliasing or non-uniform temporal spacing.
  • domain assumption Wind power at a turbine is primarily determined by wind speed magnitude and by cos(yaw misalignment) between wind direction and nacelle (Eqs. 10–11).
    This is the physical prior encoded in the geometric encoder; if effective inflow is governed by more complex aerodynamics (turbulence, wakes, blockage), the feature is incomplete. The paper's own ablation (w/o Physical Alignment) shows a small but consistent degradation.
  • domain assumption CERRA short-range forecasts issued at t0 are genuine forecasts that do not contain future observations and are operationally available at forecast time.
    Required for the claim that the model uses 'future weather forecasting' rather than reanalysis truth; §5.1 and Appendix 9.1 distinguish analysis fields for history and forecast fields for lead times, but the specific forecast product is not documented.
  • domain assumption The K nearest CERRA grid points (K=16/20/21) adequately represent the wind field over each farm; divergence/vorticity estimated by least-squares linear regression over these points (Eq. 13) is a reasonable proxy for local flow structure.
    Small K makes the linear regression brittle; if K<3 the features are set to zero. Stated in §4.2, Eqs. (13–14).
  • domain assumption Masking out timestamps with Lost Production != 0 leaves a representative test distribution; the masked metrics in §9.4 are an unbiased estimate of operational error.
    If curtailment or downtime is correlated with high winds or ramps, dropping those hours may improve apparent accuracy; the paper does not quantify the fraction of masked samples.

pith-pipeline@v1.3.0-alltime-deepseek · 21530 in / 13871 out tokens · 175890 ms · 2026-08-01T19:02:58.800475+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Fourier Geometric Wind Power Forecasting with Numerical Weather Prediction." pith.science (2026). https://pith.science/paper/ZAL2B3MT

@misc{pith2026260717095,
  author       = {Pith},
  title        = {Pith review of: Fourier Geometric Wind Power Forecasting with Numerical Weather Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZAL2B3MT}},
  note         = {Machine review of arXiv:2607.17095}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Accurate short-term wind power forecasting is essential for grid stability and operational planning, yet remains challenging due to the complex interactions between atmospheric conditions and turbine dynamics. However, existing methods fail to effectively incorporate weather forecasting with wind turbine data (i.e., SCADA), leading to suboptimal solutions. To address this, we introduce a multimodal framework that integrates historical point-based SCADA data with grid-based Numerical Weather Prediction (NWP) forecasts, which is challenging due to heterogeneous input and the complex physical wind-turbine interactions. Our approach first explicitly decomposes inputs into scalar and vector features to better capture both site-specific and geometric dependencies and then incorporates a geometric encoder to extract rotation-invariant features from wind vectors. We further leverages a Fourier Neural Operator (FNO) architecture, which performs global convolutions in the frequency domain to efficiently model long-range spatiotemporal relationships. Extensive experiments on three real-world wind farms, with weather forecasting data, demonstrate that our model consistently outperforms state-of-the-art baselines, highlighting the effectiveness of its physically-informed design. The core implementation of our method is publicly available at: https://github.com/shawn-sypiao/GWPF.

Figures

Figures reproduced from arXiv: 2607.17095 by Fan Zehui, Fugee Tsung, Hong Cheng, Jie Zhou, Juepeng Zheng, Shiyuan Piao, Yang Liu.

Figure 1
Figure 1. Figure 1: Data visualization. Gridded NWP weather forecast [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework of the proposed method. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the Geometric Encoder: (a) Wind [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Parameter-efficiency analysis across different mod [PITH_FULL_IMAGE:figures/full_fig_p009_4.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

60 extracted references · 4 canonical work pages

  1. [1]

    Abdul Fatir Ansari, Oleksandr Shchur, et al. 2025. Chronos-2: From Univariate to Universal Forecasting. arXiv:2510.15821 [cs.LG]

    Pith/arXiv arXiv 2025

  2. [2]

    Ezgi Arslan Tuncar, Şafak Sağlam, and Bülent Oral. 2024. A review of short- term wind power generation forecasting methods in recent technological trends. Energy Reports12 (2024), 197–209

    2024

  3. [3]

    Andreas Auer, Patrick Podest, Daniel Klotz, Sebastian Böck, Günter Klambauer, and Sepp Hochreiter. 2025. TiRex: Zero-Shot Forecasting Across Long and Short Horizons with Enhanced In-Context Learning. InAdvances in Neural Information Processing Systems, Vol. 38. 57529–57580

    2025

  4. [4]

    Hao Yuan Bai and Xue Liu. 2025. T-Graphormer: Using Transformers for Spa- tiotemporal Forecasting. arXiv:2501.13274 [cs.LG]

    Pith/arXiv arXiv 2025

  5. [5]

    Lei Bai, Lina Yao, Can Li, Xianzhi Wang, and Can Wang. 2020. Adaptive Graph Convolutional Recurrent Network for Traffic Forecasting. InAdvances in Neural Information Processing Systems, Vol. 33. 17804–17815

    2020

  6. [6]

    Zico Kolter, and Vladlen Koltun

    Shaojie Bai, J. Zico Kolter, and Vladlen Koltun. 2018. An Empirical Evalua- tion of Generic Convolutional and Recurrent Networks for Sequence Modeling. arXiv:1803.01271 [cs.LG]

    Pith/arXiv arXiv 2018

  7. [7]

    Bazionis, Panagiotis A

    Ioannis K. Bazionis, Panagiotis A. Karafotis, and Pavlos S. Georgilakis. 2022. A review of short-term wind power probabilistic forecasting and a taxonomy focused on input data.IET Renewable Power Generation16, 1 (2022), 77–91

    2022

  8. [8]

    Jiawei Chen and Chunhui Zhao. 2024. Addressing Spatial-Temporal Hetero- geneity: General Mixed Time Series Analysis via Latent Continuity Recovery and Alignment. InAdvances in Neural Information Processing Systems, Vol. 37. 17910–17946

    2024

  9. [9]

    Kyunghyun Cho, Bart van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation. InProceedings of the 2014 Conference on Empirical Methods in Natural Language Processing. 1724–1734

    2014

  10. [10]

    Alex Clerc and Elizabeth Lingkan. 2025. Hill of Towie wind farm open dataset. doi:10.5281/zenodo.14870023 Version 1.0.0

    doi:10.5281/zenodo.14870023 2025

  11. [11]

    Simon Daenens, Timothy Verstraeten, Pieter-Jan Daems, Ann Nowé, and Jan Helsen. 2025. Spatio-temporal graph neural networks for power prediction in offshore wind farms using SCADA data.Wind Energy Science10 (2025), 1137– 1152

    2025

  12. [12]

    Robin de Jong and Bart Doekemeijer. 2026. LES dataset of SMARTEOLE. doi:10. 5281/zenodo.18888663 Version 1.0

    2026

  13. [13]

    Xiaochong Dong, Zhihang Mao, Yingyun Sun, and Xinzhi Xu. 2024. Short-Term Wind Power Scenario Generation Based on Conditional Latent Diffusion Models. IEEE Transactions on Sustainable Energy15, 2 (2024), 1074–1085

    2024

  14. [14]

    Zheng Dong, Renhe Jiang, Haotian Gao, Hangchen Liu, Jinliang Deng, Qingsong Wen, and Xuan Song. 2024. Heterogeneity-Informed Meta-Parameter Learning for Spatiotemporal Time Series Forecasting. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 631–641

    2024

  15. [15]

    Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam

    Vijay Ekambaram, Arindam Jati, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. TSMixer: Lightweight MLP-Mixer Model for Multivari- ate Time Series Forecasting. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 459–469

    2023

  16. [16]

    Yuchen Fang, Yuxuan Liang, Bo Hui, Zezhi Shao, Liwei Deng, Xu Liu, Xinke Jiang, and Kai Zheng. 2025. Efficient Large-Scale Traffic Forecasting with Transformers: A Spatial Data Management Perspective. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 307–317

    2025

  17. [17]

    Foley, Paul G

    Aoife M. Foley, Paul G. Leahy, Antonino Marvuglia, and Eamon J. McKeogh

  18. [18]

    Yarin Gal and Zoubin Ghahramani. 2016. Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning. InProceedings of the 33rd International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 48). 1050–1059

    2016

  19. [19]

    Schoenholz, Patrick F

    Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. 2017. Neural Message Passing for Quantum Chemistry. InProceedings of the 34th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 70). 1263–1272

    2017

  20. [20]

    Tilmann Gneiting and Adrian E. Raftery. 2007. Strictly Proper Scoring Rules, Prediction, and Estimation.J. Amer. Statist. Assoc.102, 477 (2007), 359–378

    2007

  21. [21]

    Raftery, Anton H

    Tilmann Gneiting, Adrian E. Raftery, Anton H. Westveld, and Tom Goldman. 2005. Calibrated Probabilistic Forecasting Using Ensemble Model Output Statistics and Minimum CRPS Estimation.Monthly Weather Review133, 5 (2005), 1098–1118

    2005

  22. [22]

    Kurt Schaldemose Hansen, Nikola Vasiljevic, and Steen Arne Sørensen. 2022. SCADA data from Norre_m2 wind farm. doi:10.11583/DTU.19076756.v1 Accessed: 2026-06-05

    work page doi:10.11583/dtu.19076756.v1 2022

  23. [23]

    Inam Ul Haq, Abhishek Kumar, and Pramod Singh Rathore. 2025. Machine learn- ing approaches for wind power forecasting: a comprehensive review.Discover Applied Sciences7 (2025), 1139

    2025

  24. [24]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. InAdvances in Neural Information Processing Systems, Vol. 33. 6840–6851

    2020

  25. [25]

    Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long Short-Term Memory. Neural Computation9, 8 (1997), 1735–1780

    1997

  26. [26]

    2025.Electricity 2025: Analysis and forecast to 2027

    International Energy Agency (IEA). 2025.Electricity 2025: Analysis and forecast to 2027. Technical Report. International Energy Agency, Paris, France. https: //www.iea.org/reports/electricity-2025 Accessed 2026-02-02

    2025

  27. [27]

    2022.Climate Change 2022: Mitigation of Climate Change

    IPCC. 2022.Climate Change 2022: Mitigation of Climate Change. Contribution of Working Group III to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change. Technical Report. Cambridge University Press. https: //www.ipcc.ch/report/ar6/wg3/ Accessed 2026-02-02

    2022

  28. [28]

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. 2018. Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting. InInternational Conference on Learning Representations

    2018

  29. [29]

    Yang Liu, Jiashun Cheng, Haihong Zhao, Tingyang Xu, Peilin Zhao, Fugee Tsung, Jia Li, and Yu Rong. 2024. SEGNO: Generalizing Equivariant Graph Neural Networks with Physical Inductive Biases. InThe Twelfth International Conference on Learning Representations

    2024

  30. [30]

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. 2024. iTransformer: Inverted Transformers Are Effective for Time Series Forecasting. InThe Twelfth International Conference on Learning Representations

    2024

  31. [31]

    Yang Liu, Zinan Zheng, Jiashun Cheng, Fugee Tsung, Deli Zhao, Yu Rong, and Jia Li. 2025. CirT: Global Subseasonal-to-Seasonal Forecasting with Geometry- inspired Transformer. InThe Thirteenth International Conference on Learning Representations

    2025

  32. [32]

    Yang Liu, Zinan Zheng, Yu Rong, Deli Zhao, Hong Cheng, and Jia Li. 2025. Equivariant and Invariant Message Passing for Global Subseasonal-to-seasonal Forecasting. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1879–1890

    2025

  33. [33]

    Minbo Ma, Peng Xie, Fei Teng, Bin Wang, Shenggong Ji, Junbo Zhang, and Tianrui Li. 2023. HiSTGNN: Hierarchical spatio-temporal graph neural network for weather forecasting.Information Sciences648 (2023), 119580

    2023

  34. [34]

    Md Mahmuddun Nabi Murad, Mehmet Aktukmak, and Yasin Yilmaz. 2025. WP- Mixer: Efficient Multi-Resolution Mixing for Long-Term Time Series Forecasting. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 19581– 19588

    2025

  35. [35]

    Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam

    Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. In The Eleventh International Conference on Learning Representations

    2023

  36. [36]

    Charlie Plumley. 2022. Kelmarsh wind farm data. doi:10.5281/zenodo.5841834 Version 0.0.3

    work page doi:10.5281/zenodo.5841834 2022

  37. [37]

    Charlie Plumley. 2022. Penmanshiel Wind Farm Data. doi:10.5281/zenodo.5946808 Version 0.0.2

    work page doi:10.5281/zenodo.5946808 2022

  38. [38]

    Hong Qiu, Kaikai Shi, Renfang Wang, Liang Zhang, Xiufeng Liu, and Xu Cheng

  39. [39]

    Kashif Rasul, Calvin Seward, Ingmar Schuster, and Roland Vollgraf. 2021. Au- toregressive Denoising Diffusion Models for Multivariate Probabilistic Time Series Forecasting. InProceedings of the 38th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 139). 8857–8868

    2021

  40. [40]

    David Salinas, Valentin Flunkert, Jan Gasthaus, and Tim Januschowski. 2020. DeepAR: Probabilistic forecasting with autoregressive recurrent networks.Inter- national Journal of Forecasting36, 3 (2020), 1181–1191

    2020

  41. [41]

    Schimanke, M

    S. Schimanke, M. Ridal, P. Le Moigne, L. Berggren, P. Undén, R. Randria- mampianina, U. Andrea, E. Bazile, A. Bertelsen, P. Brousseau, P. Dahlgren, L. Edvinsson, A. El Said, M. Glinton, S. Hopsch, L. Isaksson, R. Mladek, E. Olsson, A. Verrelle, and Z. Q. Wang. 2021. CERRA sub-daily regional reanalysis data for Europe on height levels from 1984 to present....

    work page doi:10.24381/cds.38b394e6 2021

  42. [42]

    Zezhi Shao, Zhao Zhang, Fei Wang, Wei Wei, and Yongjun Xu. 2022. Spatial- Temporal Identity: A Simple yet Effective Baseline for Multivariate Time Series Forecasting. InProceedings of the 31st ACM International Conference on Informa- tion & Knowledge Management. 4454–4458

    2022

  43. [43]

    Zezhi Shao, Zhao Zhang, Wei Wei, Fei Wang, Yongjun Xu, Xin Cao, and Chris- tian S. Jensen. 2022. Decoupled Dynamic Spatial-Temporal Graph Neural Net- work for Traffic Forecasting.Proceedings of the VLDB Endowment15, 11 (2022), 2733–2746

    2022

  44. [44]

    Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. 2025. Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts. InThe Thirteenth International Conference on Learning Repre- sentations

    2025

  45. [45]

    Yue Song, Diyin Tang, Jinsong Yu, Zetian Yu, and Xin Li. 2023. Short-Term Fore- casting Based on Graph Convolution Networks and Multiresolution Convolution Neural Networks for Wind Power.IEEE Transactions on Industrial Informatics19, 2 (2023), 1691–1702

    2023

  46. [46]

    Ram Dyuthi Sristi, Ofir Lindenbaum, Shira Lifshitz, Maria Lavzin, Jackie Schiller, Gal Mishne, and Hadas Benisty. 2024. Contextual Feature Selection with Con- ditional Stochastic Gates. InProceedings of the 41st International Conference on Fourier Geometric Wind Power Forecasting with Numerical Weather Prediction KDD ’26, August 09–13, 2026, Jeju Island, ...

    2024

  47. [47]

    Zhang, and Jun Zhou

    Shiyu Wang, Haixu Wu, Xiaoming Shi, Tengge Hu, Huakun Luo, Lintao Ma, James Y. Zhang, and Jun Zhou. 2024. TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting. InThe Twelfth International Conference on Learning Representations

    2024

  48. [48]

    Yuxuan Wang, Haixu Wu, Jiaxiang Dong, Guo Qin, Haoran Zhang, Yong Liu, Yunzhong Qiu, Jianmin Wang, and Mingsheng Long. 2024. TimeXer: Empowering Transformers for Time Series Forecasting with Exogenous Variables. InAdvances in Neural Information Processing Systems, Vol. 37. 469–498

    2024

  49. [49]

    Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. 2023. TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis. InThe Eleventh International Conference on Learning Representations

    2023

  50. [50]

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. 2021. Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Fore- casting. InAdvances in Neural Information Processing Systems, Vol. 34. 22419– 22430

    2021

  51. [51]

    Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, Xiaojun Chang, and Chengqi Zhang. 2020. Connecting the Dots: Multivariate Time Series Forecasting with Graph Neural Networks. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 753–763

    2020

  52. [52]

    Xin Xu, Qinglong Cao, Ruizhe Deng, Zhiling Guo, Yuntian Chen, and Jinyue Yan. 2025. A cross-dataset benchmark for neural network-based wind power forecasting.Renewable Energy254 (2025), 123463

    2025

  53. [53]

    Mei Yu, Zhuo Zhang, Xuewei Li, Jian Yu, Jie Gao, Zhiqiang Liu, Bo You, Xiaoshan Zheng, and Ruiguo Yu. 2020. Superposition Graph Neural Network for offshore wind power prediction.Future Generation Computer Systems113 (2020), 145–157

    2020

  54. [54]

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. 2023. Are Transformers Effective for Time Series Forecasting?. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 11121–11128

    2023

  55. [55]

    Hanyu Zhang, Reza Zandehshahvar, Mathieu Tanneau, and Pascal Van Henten- ryck. 2025. Weather-Informed Probabilistic Forecasting and Scenario Generation in Power Systems.Applied Energy384 (2025), 125369

    2025

  56. [56]

    Zinan Zheng, Yang Liu, and Jia Li. 2025. Mesh Interpolation Graph Network for Dynamic and Spatially Irregular Global Weather Forecasting. InAdvances in Neural Information Processing Systems, Vol. 38. 25287–25311

    2025

  57. [57]

    Zinan Zheng, Yang Liu, Jia Li, Jianhua Yao, and Yu Rong. 2024. Relaxing Con- tinuous Constraints of Equivariant Graph Neural Networks for Broad Physical Dynamics Learning. InProceedings of the 30th ACM SIGKDD Conference on Knowl- edge Discovery and Data Mining. 4548–4558

    2024

  58. [58]

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. 2021. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. 11106–11115. 9 Appendices 9.1 Dataset Construction Table 8: Dataset summary and forecasting setup. F...

    2021

  59. [2012]

    Renewable Energy37, 1 (2012), 1–8

    Current methods and advances in forecasting of wind power generation. Renewable Energy37, 1 (2012), 1–8

    2012

  60. [2024]

    A novel temporal–spatial graph neural network for wind power forecasting considering blockage effects.Renewable Energy227 (2024), 120499

    2024