the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 09 Mar 2026
Modeling wind farm response: a modular, integrated, and multi-stakeholder approach
Abstract. Accurate and computationally efficient modeling of wind farm response is essential to support a wide range of stakeholders, including research institutes, wind turbine and wind farm designers, operators, and control algorithm developers. This paper presents a modular and integrated framework for modeling wind farm response, enabling consistent and multi-purpose predictions across turbine- and farm-level applications. The proposed approach combines computationally efficient and site-specific wind farm flow modeling, high-fidelity aero-servo-elastic simulations, wake-resolved inflow characterization, and data-driven response surrogates within a flexible architecture that allows individual components to be independently developed, validated, and exchanged.
Within this framework, key novelties are introduced such as a modular and holistic wind farm model, as well as a wake-slice methodology to represent local waked inflow conditions in a compact and physically meaningful form, enabling efficient training of response surrogate models using single-turbine simulations. Artificial neural network surrogates are developed to predict individual turbine responses based on a reduced set of local inflow and control descriptors, allowing the effects of wakes, turbulence, and operational strategies to be captured without resorting to full farm-level aeroelastic simulations. Another key feature of the proposed framework is its ability to consistently model multiple turbine types as well as a wide range of operational modes (power production, start-up, shut-down and parking) combined with several control modes (normal operation, yaw-steering, derating, down-regulation and noise-curtailment) within a single formulation. To this end, the methodology employs location-agnostic load surrogates, applicable to a given turbine type irrespective of its position within a farm and at any site. The overall framework is wind farm agnostic, with a modular structure that enables application to arbitrary farm layouts, environmental conditions, and operating modes without structural modification.
The framework is tested using one open-source reference turbine and two anonymized commercial turbines. For each turbine type, surrogates were developed using a single holistic library of inflow profiles representing clean and waked conditions. The performances are evaluated through an exemplary wind farm configuration composed of six turbines, demonstrating the location agnosticism of the proposed approach. Furthermore, the framework is systematically evaluated through surrogate validation and analysis across different turbine types, environmental conditions, and operational and control modes. The results demonstrate that the proposed toolchain accurately reproduces the load variations induced by wake interactions, operational modes and control modes, while maintaining a low computational cost. By combining modular physics-based modeling with scalable data-driven surrogates, the framework provides a multi-stakeholder solution for wind farm response modeling, supporting applications ranging from design analysis to operational assessment and wind farm control studies.
Competing interests: At least one of the (co-)authors is a member of the editorial board of Wind Energy Science.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.- Preprint
(3870 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on wes-2026-45', Anonymous Referee #1, 06 Apr 2026
-
RC2: 'Comment on wes-2026-45', Anonymous Referee #2, 21 Apr 2026
The paper introduces a new approach to model numerically loads and performance of wind turbines within a farm balancing physics-based models with surrogate models. The work is one of the scientific advancements achieved during thee EU-funded project TWAIN and the scientific approach is rigorously documented.
The paper deserves to be published in the journal of Wind Energy Science. This said, I have some optional recommendations for improvements for the authors.
- Readability can be improved: I spent several hours (~16?) across multiple days to digest the document. I initially started reviewing digitally, and I then had to switch to a printed copy to effectively go back and forth multiple times across different sections and figures. I personally did not enjoy that a paper of 39 pages, excluding references and appendices, is only made of four sections. Also, some sections are long and dense, and the use of subsections would greatly improve readability. Introduction is made of 4 dense pages (in page 4, wouldn’t a bulleted list be a better option to clearly highlight the novel contributions of this work?). Also, Section 3.4 is 6 pages long. Maybe break it up into subsections?
- Modularity: the paper highlights the modularity of the approach. However, the argument in support of such modularity seems weak. The paper does not clearly explain what modularity means. What are the clear interfaces between steps of the framework? To this point, a more detailed Figure 1 would greatly help readers understand the steps of the framework, the flow of data, and would help follow the discussion of the paper.
- Ability to overcome confidentiality barriers: the paper argues that the new approach helps overcome existing barriers to protect the confidentiality of the data. I find this claim overly optimistic. As you write, aeroelastic models are owned by turbine manufacturers and are seldomly shared with developers and operators (and academia). I don’t believe manufacturers treat surrogate models capable of accurately predicting performance and loads much different. Although it’s true that surrogate models offer an additional layer of protection to manufacturers against the inverse engineering of their designs compared to aeroelastic models, the claim that these surrogates will be shared with partners seems overoptimistic. Maybe a better discussion about the modularity of the new approach would also help here. Your industry co-authors can also help shed more light on these aspects.
- Computational efficiency: the paper argues that the new framework is more computationally efficient than traditional approaches. While this is not too hard to believe, the paper does not quantify the advantages. Also, previous approaches based on thousands of DWM simulations are well suited for embarrassing parallelism. So, in presence of large computational resources and optimized computational setups, the actual computational time can be quite small (few hours for the training of a surrogate).
- Figures: figures often appear much after the paragraph discussing them. In part, this issue will be resolved in the final publication, but during review this aspect was quite painful. One extreme example is in page 5, where Figure 3(e) is mentioned. Figure 3 appears 5 pages later. Also, only 3e is called out, without explaining the other subfigures.
- Page 7 line 182: this deliverable is hard to find, and the DOI seems to point to the webpage of the project Twain, and not to the actual deliverable. Also, the deliverable itself is not that useful if you are interested in the details of the implementation. Maybe point readers to your code repository? Lastly, the reader is not being helped by the fact that Figure 2 is a subset of Figure 1, but the inputs and outputs to the flowcharts are not 100% clear nor consistent.
- Guillore 2026b is not publicly available. The paper will hopefully appear in the coming weeks, but this was another burden on readability. Also, the new sampling approach described in Page 11 lines 274-280 is a clear aspect of novelty compared to Guillore 2024. The paper does not really discuss why it was introduced compared to the old, gridded approach. Maybe this is discussed in the 2026 paper? Regardless, I would highlight and discuss this aspect of novelty compared to already published work.
- Page 12 lines 316-320: I don’t fully follow why shutdown and startup cases were extended to 10 minutes. This choice does not look consistent with IEC standards. Wouldn’t it be more rigorous to include the fatigue damage caused by a pre-assumed number of startups and shutdowns over the life of each turbine?
- I would split Section 3 into four different sections: 3) Turbine models and simulation setup 4) Surrogate setup, training, and performance 5) Single turbine loads and performance 6) Wind farm loads and performance. I think different readers would be interested in different sections, and you’d help their reading by clearly separating the different topics.
- Page 26 lines 613-623: again, a proposed improvement to the readability, but this paragraph ends up masking a highlight of the present work, which starts at line 624.
- Page 28: currently, the discussion feels verbose and somewhat contorted. Unless I am missing something, the study presented in Section 3.4 and in Figures 11 and 12 could have been performed in a standard aeroelastic solver without the need of the more sophisticated approaches described here. It was a bit of a disappointment to read so many pages to then read some discussion about the impact of TI on loads, which does not seem to require fancy surrogates. Related to this aspect, I am confused about the intended takeaway of Figures 11 and 12. Figure 11 is designed to compare the effect of TI. Wouldn’t it have been more impactful to discuss how different turbine designs react differently to same levels of TI? Naively I would have split the TI levels across different subplots, where for each subplot one could immediately compare turbine A vs turbine B. I thought this would help the narrative of the paper more than a less impactful story about the impact of different levels of TI. Figure 12 raises a similar doubt. I don’t really follow what the intended takeaway here is. And why does start up stop at 12 m/s? I would guess that turbines need to be able to start up at any wind speed.
- Page 28 line 664: less important, but why “As expected”? In my experience many smaller ~2MW turbines are driven by storm loads (without opening the lid on the accuracy of standard engineering models for storm loads…).
- Page 34: wouldn’t it be better if these formulas and paragraphs appear in the methodological sections 2.x?
- Figure 17: again, I am not sure what the intended takeaway here is. Also, being picky here, but stylistically I would have put the color bar on the right of each plot, and not right below the x-axis. Currently the labels of the x-axes seem to refer to the color bar.
- Figure 18 has again a lot of data and its takeaways are not exactly immediate. If the conclusion is that wind farm control through wake steering does not negatively impact loads nor clearly benefit performance for this given wind farm for either turbine design, I think that’s a conclusion worth highlighting, and I don’t think Figure 18 clearly brings home this message.
- Conclusions are again 2.5 pages of dense text. Why not splitting them to highlight the takeaways of the work? I suspect that the paper would be easier to read, be more accessible, and ultimately become more impactful.
Thank you for letting me review your work and congratulations on a successful and impactful research project.
Citation: https://doi.org/10.5194/wes-2026-45-RC2
Data sets
Datasets as supplement for the journal publication "Modeling wind farm response: a modular, integrated, and multi-stakeholder approach" Adrien Guilloré et al. https://doi.org/10.5281/zenodo.18634031
Model code and software
Code repositories of the wind farm response framework and data preparation scripts as supplement for the journal publication "Modeling wind farm response: a modular, integrated, and multi-stakeholder approach" Andreas Vad et al. https://doi.org/10.5281/zenodo.18633503
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 219 | 92 | 19 | 330 | 17 | 16 |
- HTML: 219
- PDF: 92
- XML: 19
- Total: 330
- BibTeX: 17
- EndNote: 16
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The manuscript presents a framework for surrogate modeling of wind farms. While not providing breakthroughs compared to previous loads surrogate work, it is a good overview and can be a useful standard reference paper.
The paper lacks evidence for some of the claims in my opinion:
- The toolchain is claimed to have a low computational cost, but I cannot find that numbers were given to support the claim directly (although of course we can trust it is relatively quick compared to the models providing the training data). How quick/low computational cost are the simulations with the surrogate?
- The toolchain is claimed to be fully modular and it is said that blocks can be independently enhanced / updated (line 100-105). There is no clear evidence given imo. Further it is mentioned that wake parameters are to be calibrated in line 186-187, referring to Braunbehrens, 2023. Indeed, the assumption mentioned in line 192-194 is very critical. My own experience is that the turbine loads surrogate models need to be retrained when changing the flow model to get reliable results. Hence maybe the the first statement on independent modular model building is too broad/optimistic and we need to be very careful with changing modeling blocks (retraining/calibration may be needed)?
Furthermore in relation with the previous point, the methodology is not quite clear to me in Section 3.4. The wake model used now is a FLORIS model instead of FAST.FARM DWM as used for training the surrogate. Was any recalibration of the FLORIS wake model parameters applied? How do the FLORIS results of Figure 13 compare to FAST.FARM model results? That to me seems a very relevant verification of the framework?
Further, I think the statement in Line 70 generalizes too much in the way it is formulated, considering that many types of wind farm models are applied in industry and research. In the same paragraph, widely used turbulence models are mentioned, but we can also think of electrical modelling, environmental impact modeling etc.
Table B1 has a formatting problem (does not respect the page margins).