Pith. sign in

REVIEW 4 major objections 4 minor 63 references

Probabilistic Wind Power Forecasting with Tree-Based Machine Learning and Weather Ensembles

T0 review · 4 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Pith's one-line read This paper argues that tree-based probabilistic forecasters, led by a conditional diffusion model, cut day-ahead wind power error by up to 53% over power-curve baselines and gain further accuracy from an ensemble of weather forecasts.

desk verdict Useful empirical comparison, but the abstract promises a GP baseline that never appears and the 'best overall' claim excludes storms. read the letter →

arxiv 2602.13010 v2 pith:5DB7BWZC submitted 2026-02-13 cs.LG

classification cs.LG
keywords probabilisticforecastingwindpowergradientboostingtreesconditionaldiffusionmodelsconformalizedquantileregressionnaturalweatherensemblesoffshorefarms
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 asks which probabilistic method should turn day-ahead wind-speed forecasts into power forecasts, and it argues that tree-based machine learning paired with a conditional diffusion model is the most accurate option across Belgium's offshore wind farms. Using four years of data, the authors show mean absolute error reductions of up to 53% over the raw power curve and up to 33% over a calibrated analytical wake model, while also producing a full predictive distribution. All three tree-based probabilistic methods outperform a Gaussian-process baseline in probabilistic skill, and feeding the model an ensemble of five weather forecasts improves point accuracy by up to 23%. The practical upshot is that data-driven probabilistic forecasts can directly replace deterministic engineering baselines for day-ahead operations, at least within the wind-speed regions tested.

What carries the argument

The central object is the combination of gradient-boosting trees with a probabilistic prediction layer, where the layer that carries the paper's argument is a conditional diffusion model: a score-based diffusion process that learns the conditional distribution of power given weather features and produces 50 samples per forecast, converted to empirical quantiles. It is distribution-free, which matters because wind power conditional distributions are asymmetric near cut-in and rated wind speeds, where a Normal assumption (natural gradient boosting) yields negative power values and worse scores. The same wrapper framework also includes conformalized quantile regression, which nearly matches the

What would settle it

Evaluate the same trained models on a held-out set of storm hours—wind speed forecasts above the turbine cut-out threshold—and compare mean absolute error and continuous rank probability score per farm; if the conditional diffusion model's edge over the wake model and conformalized quantile regression disappears, then the 'best overall' claim is limited to non-storm conditions.

Watch

Extended reading notes

Core claim

The paper's central empirical claim is that conditional diffusion models trained on gradient-boosted trees give the best day-ahead probabilistic and point forecasts of wind power for all nine Belgian offshore wind farms, reducing mean absolute error by 53% versus the manufacturer power curve and by 33% versus a calibrated analytical wake model. On probabilistic skill, the diffusion model improves the continuous rank probability score by 12% over a stochastic Gaussian-process baseline, and its point forecast is 5% better in mean absolute error than that baseline. The authors further show that feeding a model an ensemble of five weather forecasts rather than a single provider improves point ac

Load-bearing premise

The paper's evaluation stops below the wind speed at which turbines shut down for safety, the storm conditions the authors identify as causing the largest day-ahead errors, so the claimed superiority has not been shown in that regime.

Editorial extensions

If this is right

  • Grid operators can expect roughly halved day-ahead mean absolute error from tree-based probabilistic forecasters compared to raw power-curve estimates, with full predictive distributions as a by-product.
  • An ensemble of weather forecasts from several providers is a practical lever: it yields up to 23% point-accuracy improvement over the worst single provider and approaches reanalysis-based accuracy.
  • The calibrated analytical wake model cuts MAE by 30% versus the power curve, so physical wake modelling is substantially closer to ML accuracy than the raw power curve and remains preferable when interpretability matters.
  • Among probabilistic wrappers, the conditional diffusion model is best but computationally heavier; conformalized quantile regression is the near-competitive alternative for operational settings.
  • Distribution choice matters: assuming a Normal shape for wind power hurts performance, especially around cut-in and rated wind speeds where the conditional distribution is strongly asymmetric.

Reading between the lines

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

  • If the same ranking holds outside the tested operating envelope, wind-farm operators could use the diffusion model's sampled distributions directly in day-ahead bidding, converting forecast skill into revenue rather than just error metrics.
  • The near-parity between ensemble-based ML and reanalysis-based ML suggests that most remaining error is the weather forecast's irreducible uncertainty; adding more diverse forecast members or higher-resolution local grids may yield further but diminishing gains.
  • Because conformalized quantile regression almost matches the diffusion model at lower cost, the economically rational industrial deployment may be CQR for high-throughput assets and diffusion for critical or high-value farms; the paper itself does not make this recommendation.
  • A natural next experiment is to re-run the comparison inside the excluded storm region: if the diffusion model's advantage persists there, its 'best overall' claim becomes robust for emergency and balancing contexts.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper presents a comparative study of three probabilistic forecasting methods—conformalized quantile regression (CQR), natural gradient boosting (NGBoost), and conditional diffusion models (Treeffuser)—all combined with gradient-boosted trees for day-ahead wind power forecasting. The models are trained on four years of data from all Belgian offshore wind farms and use an ensemble of five numerical weather prediction (NWP) forecasts. Point forecasts are also compared against two deterministic engineering baselines: a manufacturer power curve and a calibrated analytical wake model. The results in Tables I and II indicate that Treeffuser achieves the lowest average MAE and CRPS, and Figure 6 suggests that using an ensemble of NWP inputs improves accuracy relative to individual NWP providers. The abstract additionally claims comparisons against a stochastic variational Gaussian process baseline and specific improvement percentages (5% MAE, 12% CRPS, 17% average ensemble gain) that do not appear in the body.

Significance. If the reported results are robust, the study makes a useful contribution: it evaluates three modern probabilistic tree-based methods on a relatively large, real-world offshore wind dataset, includes a comparison against physically motivated wake-model baselines, and demonstrates the value of using multiple NWP ensemble members. The use of four years of data and all Belgian offshore wind farms is a strength relative to many single-farm studies. The paper is clearly written and the methodology is generally sound for the experiments actually presented. However, several claims in the abstract and conclusions are not supported by the material in the body, and the exclusion of storm-related operating region 4 limits the generality of the 'best overall' conclusion. These issues must be addressed before the paper can be recommended for publication.

major comments (4)
  1. [Abstract and Section IV] The abstract promises a stochastic variational Gaussian process baseline and states that 'all three methods outperform the Gaussian process baseline in probabilistic skill, while two out of the three also improve point forecast accuracy,' with specific improvements of 5% in MAE and 12% in CRPS. No such baseline appears in Section II, Section III, Section IV, or the appendix, and Tables I and II contain no GP results. The quoted improvement percentages also cannot be traced to the presented numbers: Treeffuser's average MAE is 8.0% vs. 8.5% for NGBoost (a 5.9% relative improvement) and 8.2% for CQR (2.4%), while the CRPS differences in Table II are smaller. This mismatch between abstract and body is a central reporting issue. The authors should either include the GP baseline and its results in the experiments or remove all GP-related claims from the abstract.
  2. [Table III and Section IV] Operating region 4, defined as wind speeds above cut-out, is excluded from the analysis. The text explicitly notes that this region is 'strongly associated with storm events and cut-out events, which are known to cause the largest errors in day-ahead wind power forecasts [51].' Consequently, the paper's headline claims—'best overall' for Treeffuser and MAE improvements of 53% vs. the power curve and 33% vs. the wake model—are demonstrated only for wind speeds below cut-out. If the relative ranking of methods changes under storm conditions, the central conclusion would not hold in exactly the regime where forecasts are most consequential. The authors should either evaluate region 4 (even with a robustness analysis using the limited available observations) or explicitly scope the abstract, results, and conclusion to non-storm operating regions.
  3. [Tables I and II] The paper reports single-point estimates of MAE and CRPS per wind farm and as an average, without error bars, confidence intervals, or significance tests. The abstract uses the word 'significantly,' and several comparisons are close (e.g., CQR 8.2% vs. Treeffuser (no tuning) 8.2% in Table I; Treeffuser 5.6% vs. CQR 5.7% in Table II), so the differences may not be statistically meaningful across the nine wind farms. Given the small number of farms, the authors should report paired tests (e.g., Wilcoxon signed-rank or bootstrap intervals) or at least per-farm standard deviations for the main comparisons.
  4. [Figure 6 and Section IV] The abstract claims an 'average improvement in point forecast accuracy of 17% by using an ensemble of weather forecasts instead of a single provider.' In the body, the only quantitative ensemble comparison is a 23% improvement over the worst-performing single provider (MétéoFrance). The average improvement over all five providers is not reported, and Figure 6 does not provide the underlying numbers. Since the 17% figure is a headline result, the authors should either report the per-provider improvements explicitly and define how the average is computed, or remove the specific percentage from the abstract.
minor comments (4)
  1. [Section III-B] The ERA-5 curtailment-filtering procedure uses the 5th percentile of each wind speed bin, with a bin size of 0.5 m/s, and the bin size is increased when a bin has fewer than 100 observations. The sensitivity of the results to these thresholds is not examined. A brief robustness check or a justification from prior literature would strengthen the preprocessing description.
  2. [Section II-D] The number of diffusion samples is fixed at 50. The sensitivity of the CRPS/MAE estimates to this choice is not discussed. A sentence explaining why 50 samples suffice would be helpful.
  3. [Table III] The operating region is 'determined based on the average wind speed forecast,' but it is not stated which provider's forecast is used or whether the same averaging is applied consistently for all models. Also, the CRPS entries for the wake model are 'NA'; this should be explained (e.g., the wake model is deterministic and therefore has no predictive distribution, although a point forecast would have a CRPS equal to its MAE).
  4. [General] No code or data availability statement is provided. Given the reproducibility value of the empirical comparison, the authors should consider making the preprocessing pipeline and model configuration available.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: the study is an empirical benchmark against external baselines; the few self-citations are for baseline calibration, not for the central claim, and the missing GP-baseline comparison in the abstract is an inconsistency rather than a circularity.

full rationale

The paper's claims are empirical comparisons, not derivations from an assumed result. It evaluates out-of-sample: 'The first three years are used to train the models, whereas the last year is used to evaluate the models' and hyperparameters are optimized on a held-out validation year ('the data of 2023 is held-out from the training set'). The engineering baselines are external: the power curve is a manufacturer relation and the wake model uses parameters 'calibrated for the Belgian-Dutch offshore cluster using SCADA data from multiple wind farms, following the procedure described in [53]'. These citations are to the authors' own prior work, but they are used to construct a baseline, not to define the central advantage. No fitted parameter is later renamed as a prediction, and no 'uniqueness theorem' or ansatz is imported from self-citations. The Treeffuser machine is an external method [8], not a self-cited result. Two manuscript issues are worth noting as non-circular limitations. First, the evaluation excludes operating region 4: 'region four is excluded from the analysis due to a lack of sufficient observations. This region is also strongly associated with storm events and cut-out events, which are known to cause the largest errors in day-ahead wind power forecasts.' This scopes the 'best overall' claim to wind speeds below cut-out, but it does not make the comparison tautological. Second, the abstract reports gains 'compared to the probabilistic baseline' using a 'stochastic variational Gaussian process regression' baseline, yet the body contains no such model, experiment, or table; this is an unsupported/omitted comparison, not a circular step. The claim that the untuned Treeffuser is dropped because of overfitting is a post-hoc model-selection concern, but it is based on observed train/test CRPS differences, not on a fitted quantity being re-reported as a prediction. Overall, the central comparison is self-contained against external benchmarks, so circularity burden is low.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entities; it applies existing methods. The free parameters are standard tuning and preprocessing choices. The main load-bearing assumptions are about data representativeness and the sufficiency of weather inputs.

free parameters (3)
  • ERA-5 curtailment filter threshold = 5th percentile of power within wind speed bin
    Chosen by hand to remove suspected economic curtailments; affects training data composition.
  • Wind speed bin size = 0.5 m/s (incremented if bin has fewer than 100 observations)
    Chosen per IEC standard; determines which observations are filtered as curtailment.
  • Number of diffusion samples = 50
    Chosen to compute empirical quantiles from Treeffuser; affects the sharpness of the predictive distribution.
assumptions (4)
  • domain assumption Weather forecasts are a sufficient input set for day-ahead wind power prediction.
    Models rely exclusively on five weather forecast providers (Section II-A), ignoring other inputs such as turbine availability or SCADA features.
  • domain assumption ERA-5-based filtering correctly distinguishes curtailment from genuine low power.
    Section III-B assumes low power in a wind speed bin is due to economic curtailment, which may also remove outages or faults.
  • domain assumption The training distribution (2021-2023) is representative of the test year (2024).
    Section III-A uses a temporal split; assumes no shift in wind farm behavior or weather patterns.
  • ad hoc to paper Excluding operating region 4 does not bias the comparison.
    Section IV excludes wind speeds above cut-out due to few observations, yet acknowledges this is the largest-error regime, so reported improvements may not generalize to storms.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Probabilistic Wind Power Forecasting with Tree-Based Machine Learning and Weather Ensembles." pith.science (2026). https://pith.science/paper/5DB7BWZC

@misc{pith2026260213010,
  author       = {Pith},
  title        = {Pith review of: Probabilistic Wind Power Forecasting with Tree-Based Machine Learning and Weather Ensembles},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5DB7BWZC}},
  note         = {Machine review of arXiv:2602.13010}
}
read the original abstract

Accurate production forecasts are essential for the integration of renewable energy sources into the power grid. This paper illustrates how to obtain probabilistic forecasts of wind power generation using gradient boosting trees and an ensemble of weather forecasts. To this end, we perform a comparative analysis across three state-of-the-art probabilistic prediction methods-conformalized quantile regression, natural gradient boosting and conditional diffusion models-all of which can be combined with tree-based machine learning. The methods are validated using four years of data for all Belgian offshore wind farms. We benchmark the models against the power curve and a calibrated wake model as well as a probabilistic method using stochastic variational Gaussian process regression. The tree-based models significantly reduce the mean absolute error in comparison to the deterministic baselines. Additionally, all three methods outperform the Gaussian process baseline in probabilistic skill, while two out of the three also improve point forecast accuracy. The conditional diffusion model attains the best performance, with improvements of 5% in mean absolute error and 12% in continuous rank probability score compared to the probabilistic baseline. Last, the results indicate an average improvement in point forecast accuracy of 17% by using an ensemble of weather forecasts instead of a single provider.

Figures

Figures reproduced from arXiv: 2602.13010 by the authors.

Figure 1
Figure 1. Location of the turbines in the Belgian offshore zone [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. High-level overview of the probabilistic forecasting methodology. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visual example of the conformal prediction procedure [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The wind farms present in the Belgian offshore zone (in [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Illustrative example of the different day-ahead forecasts [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Average mean absolute error (as a % of the installed [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 3 canonical work pages

  1. [51]

    Andrea Steiner, Carmen K ¨ohler, Isabel Metzinger, Axel Braun, Math- ias Zirkelbach, Dominique Ernst, Peter Tran, and Bodo Ritter. 2017. Critical weather situations for renewable energies–part a: cyclone detection for wind power.Renewable Energy, 101, 41–50

  2. [1]

    Mehrnaz Ahmadi, Hamed Aly, and Mehdi Khashei. 2025. Enhancing Power Grid Stability with a Hybrid Framework for Wind Power Forecasting: Integrating Kalman Filtering, Deep Residual Learning, and Bidirectional LSTM.Energy, 137752, 137752

  3. [2]

    Myles R Allen and David A Stainforth. 2002. Towards objective probabalistic climate forecasting.Nature, 419, 6903, 228–228

  4. [3]

    Stijn Ally, Timothy Verstraeten, Pieter-Jan Daems, Ann Now ´e, and Jan Helsen. 2025. Modular deep learning approach for wind farm power forecasting and wake loss prediction.Wind Energy Science, 10, 4, 779– 812

  5. [4]

    Jos ´e R Andrade and Ricardo J Bessa. 2017. Improving renewable energy forecasting with a grid of numerical weather predictions.IEEE Transactions on Sustainable Energy, 8, 4, 1571–1580

  6. [5]

    Anastasios N Angelopoulos and Stephen Bates. 2021. A gentle in- troduction to conformal prediction and distribution-free uncertainty quantification.arXiv preprint arXiv:2107.07511

  7. [6]

    Georgios Batzolis, Jan Stanczuk, Carola-Bibiane Sch ¨onlieb, and Chris- tian Etmann. 2021. Conditional image generation with score-based diffusion models.arXiv preprint arXiv:2111.13606

  8. [7]

    Kevin Bellinguer, Valentin Mahler, Simon Camal, and Georges Karin- iotakis. 2020. Probabilistic forecasting of regional wind power gener- ation for the eem20 competition: a physics-oriented machine learning approach. In2020 17th International Conference on the European Energy Market (EEM). IEEE, 1–6

Show all 63 references
  1. [8]

    Nicolas Beltran Velez, Alessandro A Grande, Achille Nazaret, Alp Kucukelbir, and David Blei. 2024. Treeffuser: probabilistic prediction via conditional diffusions with gradient-boosted trees.Advances in Neural Information Processing Systems, 37, 118296–118325

  2. [9]

    Diederik van Binsbergen, Pieter-Jan Daems, Timothy Verstraeten, Amir Nejad, and Jan Helsen. 2024. Scalable scada-based calibration for analytical wake models across an offshore cluster.Journal of Physics: Conference Series, 2745, 1, (Apr. 2024), 012014. doi:10.1088/1742-6 596/...

  3. [10]

    Vadim Borisov, Tobias Leemann, Kathrin Seßler, Johannes Haug, Martin Pawelczyk, and Gjergji Kasneci. 2022. Deep neural networks and tabular data: a survey.IEEE transactions on neural networks and learning systems, 35, 6, 7499–7519

  4. [11]

    Jethro Browell, Sebastian Haglund, Henrik K ¨alvegren, Edoardo Simioni, Ricardo Bessa, Yi Wang, and Dennis van der Meer. 2023. Hybrid energy forecasting and trading competition. (2023). doi:10.21 227/5hn0-8091

  5. [12]

    Max Bruninx, Timothy Verstraeten, Jalal Kazempour, and Jan Helsen

  6. [13]

    Victor Chernozhukov, Iv ´an Fern´andez-Val, and Alfred Galichon. 2010. Quantile and probability curves without crossing.Econometrica, 78, 3, 1093–1125

  7. [14]

    2018.Wind energy design

    Thomas C Corke and Robert C Nelson. 2018.Wind energy design. CRC Press

  8. [15]

    Crespo and J

    A. Crespo and J. Hernandez. 1996. Turbulence characteristics in wind- turbine wakes.Journal of Wind Engineering and Industrial Aerody- namics, 61, 1, (June 1996), 71–85. doi:10.1016/0167-6105(95)00033 -x

  9. [16]

    Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat GANs on image synthesis.Advances in neural information processing systems, 34, 8780–8794

  10. [17]

    Tony Duan, Avati Anand, Daisy Yi Ding, Khanh K Thai, Sanjay Basu, Andrew Ng, and Alejandro Schuler. 2020. Ngboost: natural gradient boosting for probabilistic prediction. InInternational conference on machine learning. PMLR, 2690–2700

  11. [18]

    Gregor Giebel and George Kariniotakis. 2017. Wind power forecast- ing—a review of the state of the art.Renewable energy forecasting, 59–109

  12. [19]

    Tilmann Gneiting, Fadoua Balabdaoui, and Adrian E Raftery. 2007. Probabilistic forecasts, calibration and sharpness.Journal of the Royal Statistical Society Series B: Statistical Methodology, 69, 2, 243–268

  13. [20]

    Francisco Gonz ´alez-Longatt, Peter Wall, and Vladimir Terzija. 2012. Wake effect in wind farm performance: steady-state and dynamic behavior.Renewable Energy, 39, 1, 329–338

  14. [21]

    L ´eo Grinsztajn, Edouard Oyallon, and Ga ¨el Varoquaux. 2022. Why do tree-based models still outperform deep learning on typical tabular data?Advances in neural information processing systems, 35, 507–520

  15. [22]

    Yannick Heiser, Farzaneh Pourahmadi, and Jalal Kazempour. 2025. Betting vs. trading: learning a linear decision policy for selling wind power and hydrogen.Sustainable Energy, Grids and Networks, 101848

  16. [23]

    Philip J Heptonstall and Robert JK Gross. 2021. A systematic review of the costs and impacts of integrating variable renewables into power grids.Nature Energy, 6, 1, 72–83

  17. [24]

    Hans Hersbach. 2000. Decomposition of the continuous ranked proba- bility score for ensemble prediction systems.Weather and F orecasting, 15, 5, 559–570

  18. [25]

    Jonathan Ho and Tim Salimans. 2022. Classifier-free diffusion guid- ance.arXiv preprint arXiv:2207.12598

  19. [26]

    Tao Hong, Pierre Pinson, Shu Fan, Hamidreza Zareipour, Alberto Troccoli, and Rob J. Hyndman. 2016. Probabilistic energy forecasting: global energy forecasting competition 2014 and beyond.International Journal of F orecasting, 32, 3, 896–913. doi:https://doi.org/10.1016/j.i jfo...

  20. [27]

    Jianming Hu, Qingxi Luo, Jingwei Tang, Jiani Heng, and Yuwen Deng. 2022. Conformalized temporal convolutional quantile regression networks for wind power interval forecasting.Energy, 248, 123497

  21. [28]

    IEA. 2025. Global energy review 2025. Licence: CC BY 4.0. Paris, (2025). https://www.iea.org/reports/global-energy-review-2025

  22. [29]

    Ivan Kati ´c, Jørgen Højstrup, and Niels Jensen. 1987. A simple model for cluster efficiency. In

  23. [30]

    Yuko Kato, David MJ Tax, and Marco Loog. 2023. A review of non- conformity measures for conformal prediction in regression.Conformal and Probabilistic Prediction with Applications, 369–383

  24. [31]

    Roger Koenker and Gilbert Bassett Jr. 1978. Regression quantiles. Econometrica: journal of the Econometric Society, 33–50

  25. [32]

    Matthias Lange. 2005. On the uncertainty of wind power predic- tions—analysis of the forecast accuracy and statistical distribution of errors.J. Sol. Energy Eng., 127, 2, 177–184

  26. [33]

    Yonggang Li, Yue Wang, and Binyuan Wu. 2020. Short-term direct probability prediction model of wind power based on improved natural gradient boosting.Energies, 13, 18, 4629

  27. [34]

    2009.Directional statistics

    Kanti V Mardia and Peter E Jupp. 2009.Directional statistics. John Wiley & Sons

  28. [35]

    Llew Mason, Jonathan Baxter, Peter Bartlett, and Marcus Frean

  29. [36]

    James E Matheson and Robert L Winkler. 1976. Scoring rules for continuous probability distributions.Management science, 22, 10, 1087–1096

  30. [37]

    Amin Niayifar and Fernando Port ´e-Agel. 2015. A new analytical model for wind farm power prediction. InJournal of physics: conference seriesnumber 1. V ol. 625. IOP Publishing, 012039

  31. [38]

    Jakub Nowotarski and Rafał Weron. 2018. Recent advances in electric- ity price forecasting: a review of probabilistic forecasting.Renewable and Sustainable Energy Reviews, 81, 1548–1568

  32. [39]

    European Network of Transmission System Operators for Electricity

  33. [40]

    Pedersen

    Mads M. Pedersen. 2023. Pywake 2.5.0: an open-source wind farm simulation tool, (Feb. 2023). https://gitlab.windenergy.dtu.dk/TOPFA RM/PyWake

  34. [41]

    Pierre Pinson. 2013. Wind energy: forecasting challenges for its operational management.Statistical Science, 28, 4, 564–585

  35. [42]

    Pierre Pinson, Christophe Chevallier, and George N Kariniotakis. 2007. Trading wind generation from short-term probabilistic forecasts of wind power.IEEE transactions on Power Systems, 22, 3, 1148–1156

  36. [43]

    Yaniv Romano, Evan Patterson, and Emmanuel Candes. 2019. Confor- malized quantile regression.Advances in neural information processing systems, 32

  37. [44]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. 2022. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10684–10695

  38. [45]

    Elia Transmission Belgium SA/NV. 2025. Elia open data platform. https://opendata.elia.be/. (2025)

  39. [46]

    Yves-Marie Saint-Drenan et al. 2020. A parametric model for wind turbine power curves incorporating environmental conditions.Renew- able Energy, 157, 754–768

  40. [47]

    Antonio Segalini and Jan- ˚Ake Dahlberg. 2020. Blockage effects in wind farms.Wind Energy, 23, 2, 120–128

  41. [48]

    Geert Smet, Joris Van den Bergh, and Piet Termonia. 2023. Probabilis- tic storm forecasts for wind farms in the north sea

  42. [49]

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. 2021. Score-based generative modeling through stochastic differential equations. InInternational Conference on Learning Representations. https://openreview.net/fo rum?id=PxTIG12RRHS

  43. [50]

    Olivier Sprangers, Sebastian Schelter, and Maarten de Rijke. 2021. Probabilistic gradient boosting machines for large-scale probabilistic regression. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, 1510–1520

  44. [52]

    Conor Sweeney, Ricardo J Bessa, Jethro Browell, and Pierre Pinson

  45. [53]

    Nejad, and Jan Helsen

    Diederik van Binsbergen, Pieter-Jan Daems, Timothy Verstraeten, Amir R. Nejad, and Jan Helsen. 2024. Hyperparameter tuning frame- work for calibrating analytical wake models using scada data of an offshore wind farm.Wind Energy Science, 9, 7, (July 2024), 1507–1526. doi:10.519...

  46. [54]

    Diederik Van Binsbergen, Pieter-Jan Daems, Timothy Verstraeten, Amir Nejad, and Jan Helsen. 2024. Performance comparison of analyt- ical wake models calibrated on a large offshore wind cluster. InJournal of Physics: Conference Seriesnumber 9. V ol. 2767. IOP Publishing, 092059

  47. [55]

    Wei Wang, Bin Feng, Gang Huang, Chuangxin Guo, Wenlong Liao, and Zhe Chen. 2023. Conformal asymmetric multi-quantile generative transformer for day-ahead wind power interval prediction.Applied Energy, 333, 120634

  48. [56]

    Paige Weber and Matt Woerman. 2024. Intermittency or uncertainty? impacts of renewable energy in electricity markets.Journal of the Asso- ciation of Environmental and Resource Economists, 11, 6, 1351–1385

  49. [57]

    Wind energy generation systems — part 1: design requirements

    2019. Wind energy generation systems — part 1: design requirements. International Standard. (Feb. 8, 2019). Retrieved May 6, 2025 from https://webstore.iec.ch/publication/26423

  50. [58]

    Siyi Zhang, Mingbo Liu, Min Xie, and Shunjiang Lin. 2024. Two-stage short-term wind power probabilistic prediction using natural gradient boosting combined with neural network.Applied Soft Computing, 159, 111669

  51. [59]

    Xinxin Zhu and Marc G Genton. 2012. Short-term wind speed fore- casting for power system operations.International Statistical Review, 80, 1, 2–23. APPENDIX The set of hyperparameters selected for model tuning is presented in Table IV. It should be noted that the natural gradie...

  52. [1999]

    Boosting algorithms as gradient descent.Advances in neural information processing systems, 12

  53. [2020]

    The future of forecasting for renewable energy.Wiley Interdis- ciplinary Reviews: Energy and Environment, 9, 2, e365

  54. [2024]

    https://transparency.entsoe.eu/

    ENTSO-E transparency platform. https://transparency.entsoe.eu/. (2024)

  55. [2025]

    InProceedings of the 16th ACM International Conference on Future and Sustainable Energy Systems, 719–726

    Day-ahead bidding strategies for wind farm operators under a one-price balancing scheme. InProceedings of the 16th ACM International Conference on Future and Sustainable Energy Systems, 719–726

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.