the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 23 Oct 2026
Design and implementation of a location-agnostic and control-oriented load surrogate model for wind turbines
Abstract. Accurate estimation of wind turbine fatigue loads is essential for wind farm design, optimization, and control applications. Load surrogate models have recently emerged as a promising approach to reduce the computational cost associated with high-fidelity load calculations. However, while previous studies have primarily focused on demonstrating predictive accuracy, considerably less attention has been devoted to understanding how such surrogates should be designed and deployed in practice.
This study systematically investigates the design and implementation of location-agnostic wind turbine fatigue load surrogate models. A comprehensive database of aeroelastic simulations covering free-stream, partially waked, and fully waked operating conditions is used to evaluate alternative inflow representations, control descriptors, number of turbulent realizations, and training dataset sizes. The practical deployment of the resulting surrogate formulation is further assessed using multiple inflow-estimation approaches, including observer-based inflow reconstruction and steady-state wake models.
A key finding of this work is that there exists a finite level of physical complexity sufficient to achieve converged performance for a general wind turbine load surrogate model. Among the investigated options, an inflow representation consisting of four sector-averaged wind speed quantities and one rotor-averaged turbulence parameter (denoted SAWS 4A + RATI) emerges as the simplest formulation capable of reproducing converged predictive accuracy. Likewise, steady-state control setpoints are shown to provide a sufficient description of control actions, approximately 10–12 turbulent realizations are required for robust convergence with respect to turbulence variability, and surrogate performance converges with a relatively modest number of training cases.
The results further demonstrate that the same design recommendations remain valid across all investigated implementation environments. Although significant performance degradation occurs when training and deployment rely on different inflow-description spaces, high predictive accuracy is consistently recovered when the surrogate is trained and deployed using the same inflow representation. These findings provide practical design guidelines for the development of fatigue-load surrogate models and demonstrate that accurate, robust, and computationally efficient load prediction can be achieved using a small set of physically meaningful reduced-order descriptors.
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
(7799 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 20 Nov 2026)
- CC1: 'Comment on wes-2026-191', J. Gordon Leishman, 24 Oct 2026 reply
-
RC1: 'Comment on wes-2026-191', Anonymous Referee #1, 03 Nov 2026
reply
Dear Authors,
I was invited to review the manuscript of the article and I went through the text to provide my comments. Overall, I think the topic of the article is timely, given the recent increase in interest within the wind energy community in including turbine loads more accurately in plant design and operation. The article represents a meaningful advancement in this area that is of interest to the scientific and industrial communities; therefore I think it deserves to be considered for publication in Wind Energy Science.
While reading the article, I had a number of comments, which are reported below. I would ask the Authors to answer them during the discussion and consider them if the article proceeds to the revision phase.General comments
The main general comment is that the article, at 75 pages, is quite long. Throughout the text, relevant information is provided and there is not a single section that can be left out to shorten the article. However, I think that concepts and sometimes entire sentences are repeated throughout the article and there is room to streamline the text by addressing this issue.In general, the abstract and introduction are more difficult to read than the remainder of the article. I think this is because they use jargon-heavy terms early on and explain the jargon and hence why the work matters too late, towards the end of the introduction. Moreover, the introduction, the motivation and the structure of the work presented in these first 6 pages can be greatly streamlined and made more compact.
Another general comment is that some conclusions that the authors generalize are specific to use cases they have in mind but are not general, so introducing the related complexity may not be of interest to everyone: 1) sector-averaged inflow is useful for predicting blade loads but not for the other load channels; 2) introducing derating and yaw setpoints is useful for blade loads and when collective control is used in the farm, which is often not the case; 3) the combination of the surrogate with a steady-state flow model is only one use case.
Specific comments
01. Page 1. Referenced text: “location-agnostic”
Comment: I think it is useful to explain here in the abstract what "location-agnostic" means using plain words.02. Page 1. After “models” in the sentence “This study systematically investigates the design and implementation of location-agnostic wind turbine fatigue load surrogate models.”
Comment: Here, I think it could be useful to briefly explain how such surrogate models work.03. Page 1. Referenced text: “control descriptors,”
Comment: “Clarify what you mean by "control descriptors".”04. Page 2. Referenced text: “site-specific inflow descriptors,”
Comment: “What are "site-specific inflow descriptors"?”05. Page 2. Referenced text: “suitable for wind farm control applications.”
Comment: “What additional requirements are posed by suitability for wind farm control applications beyond accuracy, robustness or computational efficiency?”06. Page 3. Referenced text: “generic”
Comment: “What do you mean by "generic"?”07. Page 3. Referenced text: “and investigates its compatibility with both engineering wake models and observer-based inflow estimates.”
Comment: “This seems the same as the first research question. I think the third research question can be streamlined to highlight its complementarity to the other two while avoiding repeated content.”08. Page 3. Referenced text: “Here, location-agnostic refers to a formulation that depends only on the local turbine inflow rather than on farm-specific quantities such as turbine spacing, layout geometry, or wake overlap metrics. Control-oriented refers to a representation that remains applicable under a broad range of turbine- and wind-farm-level control actions.”
Comment: “I suggest reporting these definitions earlier in the text, especially in the abstract.”09. Page 4. Referenced text: “practical estimability.”
Comment: “What is "practical estimability"? I think it is explained in the next sentence, but it is not clear.”10. Page 4. Referenced text: “curl-wake effects.”
Comment: “Explain after this sentence what a curled wake is, why it arises and why it could be meaningful to the scope of surrogate load models in wind farms.”11. Page 4. Referenced text: “this modeling framework. Future investigations may revisit these findings using wake models that incorporate explicit curl-wake dynamics.”
Comment: “This is not clear. Clarify why it would be beneficial to include curl-wake effects in future studies.”12. Page 4. Referenced text: “Near-wake effects are not explicitly assessed because they are not represented by the wake modeling framework used to generate the datasets. Future studies involving enhanced wake models and large eddy simulations could provide further insight into whether a finer spatial representation of the rotor inflow becomes necessary in the presence of strong near-wake flow structures.”
Comment: “Here it is not clear what near-wake effects could be meaningful in this framework and why. DWM models are made to capture far-wake dynamics while neglecting near-wake dynamics, and this is sufficient for predicting loads in wind farms because wind turbines normally operate in the far wake of other turbines, with sufficient distances between turbines to provide the required separation. Clarify your point with respect to this comment.”13. Page 4. Referenced text: “as a design problem.”
Comment: “Clarify what "design problem" means. What is a design problem, and how does this differ from approaches used by others?”14. Page 5. Figure 1.
Comment: “I think it is not clear why the effects of control are among "reduced-order inputs" and not among "inputs" in the implementation.”15. Page 10. Referenced text: “centered on the vertical axis” and comparison between configurations 4A and 4B shown in Figure 2.
Comment: “"centered on the vertical axis" does not clearly convey the difference from 4B.”16. Page 11. Referenced text: “RAVS, is derived by assuming a classical power-law representation of the vertical velocity profile,”
Comment: “How does this apply to a wake-related inflow where the gradient in the vertical direction is due to the combination of sheared inflow and wake deficit?”17. Page 12. Referenced text: “rather than measured directly.”
Comment: “In this case, I think that reduced-order quantities are estimated directly from wind observers. In this case, it could be good to explain this to clarify that it is different from estimating the inflow and then the reduced-order quantities from it.”18. Page 14. Referenced text: “In fact, although it might be tempting to train based on the higher fidelity inflows, at usage time, when the model can only be fed with the lower fidelity inflows, any discrepency between the two will invariably degrade the quality and correctness of the retrieved DELs.”
Comment: “How can one apply this training approach to field data? In this case, I imagine one can easily generate the large database of DELs from high-fidelity models, but how can one perform the training based on inflow descriptors coming from the field? How can one obtain inflow descriptors corresponding to the cases simulated in the high-fidelity models if the database of conditions is large?”19. Page 17. Referenced text: “The data-driven surrogate models introduced in Sect. 2.5 require a training dataset covering a broad range of inflow conditions, wake interactions, and wind-farm control actions.”
Comment: “The article explains in Sect. 3 that the present work focuses on simulation data. However, throughout the article, it is said that the formulation is generic and can be applied to different data sources such as wind tunnel and field measurements. However, datasets coming from measurements typically do not cover broad ranges of inflow and operating conditions, at least not to the same extent as numerical datasets. Is this a concern for the methodology proposed in this article? If yes, how can it be addressed, at least in principle, to maintain the general validity of the formulation, as claimed by the authors?”20. Page 18. Referenced text: “increasing the number of turbines in the training farm would not necessarily introduce fundamentally new input-output relationships for the surrogate to learn.”
Comment: “Why? Doesn't this statement imply that deep array effects are small or negligible? Clarify.”21. Page 18. Referenced text: “requires tens of thousands of simulation cases,”
Comment: “Coming back to my point on the generality of this framework: how do you expect tens of thousands of different conditions to be available in measurements? And how can this lack of data (if real) be connected to the training of the inflow–>loads mapping?”22. Page 18. Referenced text: “surrogate-training datasets.”
Comment: “Yes, but then you say that the mapping has to be learned based on the inflow data source that will be used for prediction and so, in the case of field measurements, based on inflow measurements.”23. Page 19. Referenced text: “The default DWM model parameters provided with FAST.Farm are used throughout this work.”
Comment: “Report the FAST.Farm version you are using and the repository with standard input files.”24. Page 24. Referenced text: “aerodynamic power”
Comment: “Why aerodynamic power and not electrical power at the generator? The latter is a measurable quantity; the former is not.”25. Page 26. Referenced text: “The resulting observations are used to derive practical recommendations for the development of accurate and computationally efficient load surrogate models.”
Comment: “This is repeated at the beginning of the text in Sect. 4. In general, in the article, there are sentences that are repeated a few times. Repetition can be useful, but here it seems excessive. The article is also quite long (75 pages), and I think there is room to streamline it.”26. Page 27. Referenced text: “Overall, the results in Fig. 7 reveal a clear and consistent trend across all considered load channels: prediction errors decrease as additional information describing the rotor-scale inflow field is incorporated into the surrogate inputs.”
Comment: “I would qualify this statement better. A big jump in accuracy is obtained when moving from RAWS to RAWS+RATI (except for blade-root DEL). Then other additions do not lead to equally large improvements. SAWS seems worse than RAWS (except for blade-root DEL). So, I would argue it is not true that "more information = better accuracy".”27. Page 28. Referenced text: “Overall, SAWS 4A + RATI emerges as the most attractive inflow representation among the investigated alternatives.”
Comment: “This statement also needs better qualification. For all load channels except blades, RAWS+RATI gives similar performance with lower complexity and looks more attractive.”28. Page 32. Referenced text: “This choice also simplifies the presentation of the subsequent analyses.”
Comment: “You have to either explain why it simplifies the presentation of the subsequent analyses or skip this sentence.”29. Page 32. Referenced text: “The question addressed in this subsection is which control information is required to accurately capture the influence of control actions on DELs while maintaining a compact surrogate formulation.”
Comment: “Move this sentence to the beginning of the section.”30. Page 32. Referenced text: “In contrast to the analyses presented previously, the same inflow conditions may now correspond to multiple turbine operating states and therefore to substantially different DELs. Although the inflow descriptor SAWS 4A + RATI provides an accurate representation of the local flow conditions, it does not contain sufficient information to distinguish between the operating states associated with different yaw and curtailment settings.”
Comment: “It is not clear why there are multiple operating states. I think an example (or a deeper explanation) could be useful to clarify this.”31. Page 34. Referenced text: “Once both yaw misalignment and power demand are included, most of the ambiguity introduced by the controlled operating conditions is resolved,”
Comment: “Related to my previous comment, this conclusion is useful only for cases where farm flow control is included (via derating or wake steering). I think it is important to clarify that the ambiguity the authors are talking about is present only when these control strategies are used and does not need to be resolved in other cases.”32. Page 34. Referenced text: “Further enrichment of the control vector through the addition of rotor speed, blade pitch, rotor torque, and aerodynamic power statistics produces little further reduction in NRMSE, indicating that the surrogate performance has largely reached a plateau.”
Comment: “These are the standard turbine-level control inputs, so I think a good conclusion is that, in the vast majority of cases, it is possible to get accurate load predictions by including inflow information but not control setpoints.”33. Page 34. Referenced text: “From a practical perspective, this result is particularly noteworthy because obtaining quantities such as rotor speed, blade pitch, rotor torque, aerodynamic power, and especially their temporal standard deviations would require substantially more information than is typically available from reduced-order steady-state flow models.”
Comment: “I think this sentence applies only to the specific case where the surrogate models are used in combination with steady-state models. In all other cases (aero-elastic simulations with DWM or LES, wind tunnel or field measurements), it is possible to get signals of turbine-level control.”34. Page 35. Referenced text: “In practical applications,”
Comment: “In some practical applications”35. Page 36. Referenced text: “are not always available in all reduced-order steady-state inflow models.”
Comment: “Again, to increase the appeal of the article, I think it is beneficial to avoid referring to specific use cases. Getting SATI would be difficult in any application involving real measurements (wind tunnel, field), but it is possible in DWM / LES.”36. Page 36. Referenced text: “A possible explanation is that wake meandering, which predominantly develops in the lateral direction, introduces inflow fluctuations that are more effectively captured by a left-right partition of the rotor than by an up-down partition.”
Comment: “This comment makes sense only if meandering introduces an offset in the average position of the wake over the observation window. Is this the case?”37. Page 38. Referenced text: “Along the horizontal axis, the yaw-misalignment angles of WT1, WT2, and WT3 are varied simultaneously.”
Comment: “To me, it is not clear what it means that yaw angles are varied simultaneously. Does it mean that the three turbines have the same yaw angle?”38. Page 46. Referenced text: “considerably more valuable than simply providing a larger number of examples associated with an inadequate representation.”
Comment: “Isn't this kind of obvious? Are there surrogate models where this behavior is different and providing just a larger number of examples without taking care to cover the underlying physics adequately is beneficial?”39. Page 47. Referenced text: “by first reconstructing local inflow time series and subsequently applying the corresponding temporal and spatial averaging operations.”
Comment: “Is this possible with LIDAR compared to blade load measurements?”40. Page 47. Referenced text: “Fig. D1 in Appendix D, while the corresponding comparison for the effective quantities, namely REWS, SEWS 4A, RETI, and SETI 4A, is shown in Fig. D2.”
Comment: “Since these results are discussed here, I think it is better to report the figures here instead of in the appendix.”41. Page 52. Referenced text: “Similarly to the observer-based implementation, the comparisons in Fig. E1 and Fig. E2 show that the reduced-order inflow quantities do not coincide exactly, particularly for TI.”
Comment: “Since the results are discussed here, I think that it may be worth reporting the figures here instead of in the appendix.”42. Page 57. Referenced text: “what level of physical complexity”
Comment: “This needs to be qualified better. The physical complexity analysed in the work is related to inflow and turbine control, while the rest of the modeling (servo-aero-structural dynamics) is not addressed here.”Technical corrections
43. Page 1. Add “inflow” between “turbulent” and “realizations” in the sentence “conditions is used to evaluate alternative inflow representations, control descriptors, number of turbulent realizations, and training dataset sizes.”44. Page 2. Replace “extensive literature” with “extensive literature search” (?).
45. Page 4. Replace “atmospheric” with “wind”
46. Page 9. Replace “capture” with “provide information about”
47. Page 11. Replace “identify” with “provide”
48. Page 12. Replace “dynamic models” with “models of turbine structural dynamics and aerodynamics” (?).
49. Page 26. Remove “surrogate” at line 657.
50. Page 57. Replace “turbulent” at line 1381 with “turbulent wind field”.
Citation: https://doi.org/10.5194/wes-2026-191-RC1 -
RC2: 'Comment on wes-2026-191', Anonymous Referee #2, 04 Nov 2026
reply
Dear authors, I enjoyed reading your manuscript. It presents an extensive investigation into reduced-order surrogate models for structural load estimation in wind farms. The topic is relevant, and the paper provides a systematic assessment of several factors influencing surrogate model design, including inflow representation, control variables, turbulent realisations and training-to-deployment strategies.
I particularly appreciate the effort to connect the numerical results with their physical implications. The proposed reduction of turbine inflow into sector-averaged wind speeds and rotor-averaged turbulence intensity is a valuable contribution towards computationally efficient load estimation. The investigation also provides useful practical insights into the design and implementation of load surrogates.
However, I have some concerns regarding the interpretation and generality of the conclusions.
Specific comments
1. Generalisation and simulation framework
The training and testing databases are generated using essentially the same aeroelastic and wake simulation framework. Although testing on a different farm layout provides a useful assessment of transferability across turbine locations, it does not establish generalisation beyond the underlying simulation models and assumptions. The surrogate is effectively learning the load relationships represented by the simulation framework, including its limitations. I would encourage the authors to distinguish more clearly between location-agnostic predictions and generalisation across different turbine models, controllers, wake models or real operating conditions.
2. Inflow representation
The investigation of sector-based inflow descriptors is valuable, particularly the combination of SAWS 4A and RATI. However, the effectiveness of these representations appears to depend on the structural load channel being investigated. I would encourage the authors to clarify the scope of the proposed formulation and distinguish the novelty of the present investigation from the sector-based approaches developed in their previous work.
3. Representation of control actions
The conclusion that power demand provides sufficient information, while additional turbine-state variables offer limited improvements, deserves further discussion. In particular, pitch control directly influences aerodynamic loading and structural load variability. Given that the database is concentrated around 8–12 m/s, I wonder whether the investigated operating conditions sufficiently capture the influence of pitch regulation, particularly around rated operation. The conclusion regarding the sufficiency of power demand may therefore be specific to the investigated operating range and control strategy rather than generally applicable to load surrogate modelling.
4. Number of turbulent realisations
The identification of 10–12 turbulent realisations as a good compromise between predictive accuracy and computational cost is an interesting practical finding. However, this conclusion appears predominantly empirical, based on the observed NRMSE trends. I would have expected a stronger statistical justification, particularly considering the extensive simulation campaign. It would be useful to distinguish between the number of realisations required to achieve stable surrogate prediction accuracy and the number required to obtain statistically representative structural load responses.
5. Training and validation strategy
The random shuffling followed by five-fold cross-validation does not necessarily guarantee independent training and validation datasets. Samples originating from the same turbulent realisation or operating condition may be distributed across different folds, potentially preserving correlations between training and validation data. Have the authors considered grouping samples by turbulent realisation or simulation condition to provide a more robust assessment of generalisation to unseen turbulent conditions?
Final assessment
Overall, I consider this a valuable and technically interesting investigation. My main concerns relate to the extent to which the findings and recommendations can be generalised beyond the particular simulation framework and operating conditions considered. Clarifying these limitations and strengthening the justification of some of the conclusions would improve the manuscript and better establish the applicability of the proposed surrogate modelling approach.
Citation: https://doi.org/10.5194/wes-2026-191-RC2 -
RC3: 'Comment on wes-2026-191', Anonymous Referee #3, 13 Nov 2026
reply
The manuscript presents a surrogate modelling framework for predicting wind turbine fatigue loads from inflow information and control parameters, focusing on how to identify a compact set of inputs that maintains sufficient predictive accuracy. An extensive numerical database is used for training and validation, and performance is assessed using an independent testing database. The systematic comparison of input representations provides useful insights into the balance between model complexity and prediction accuracy. The authors relate their findings to the underlying physical behaviour and offer practical guidance for applying and further developing the framework.
In my opinion, this is a valuable contribution to wind turbine load surrogate modelling. The manuscript is generally well written and provides thoughtful discussion of the model, the influence of its inputs, and the underlying physics.
One general comment is that readers would benefit from a more concise presentation of the key recommendations, as some practical guidance is embedded in lengthy paragraphs. For example, guidance on input selection is mainly presented in discussions, while application to different turbine models and control strategies, which can be a key issue, is later addressed in the conclusions as future work. If feasible, a short, dedicated section summarising the recommendations and limitations for practical use could improve readability and help readers apply the findings.
The following comments concern clarification and possible further discussion.
1. The comparison of sectorisation strategies A and B would benefit from examining whether their similar aggregate performance conceals differences associated with particular inflow configurations.
Figure 5 suggests that changing the sector orientation can produce appreciable differences in the quantities assigned to individual sectors. Even with the same number of sectors, localised regions of non-uniform wind speed or elevated turbulence may be represented differently depending on their position on the rotor disk. Strategies A and B could therefore perform differently for specific inflow configurations despite yielding similar overall NRMSE values and predicted-load distributions.
For example, one strategy could systematically overpredict loads for one subset of conditions and underpredict them for another. Although these errors would not cancel in the squared-error calculation, their direction and association with specific conditions may remain unclear from aggregate NRMSE alone.
Extending the analysis in Figure 12 to include the four-sector B configuration would be helpful. Comparing prediction errors across yaw misalignment conditions could reveal any consistent bias associated with sector orientation or inflow asymmetry. A representative comparison using the existing database could strengthen the discussion of when the recommended sectorisation is sufficient and whether particular conditions warrant finer resolution.
2. Clarifying the definition and implementation of “power demand” would help readers interpret the contribution of the control-related inputs.
Appendix B indicates that power demand is prescribed as part of the operating strategy. Its effect on loads is mediated by the implemented controller and may involve coordinated changes in rotor speed, blade pitch, and generator torque. Power demand may therefore already provide substantial information about the operating state within the investigated dataset.
This could explain the limited additional improvement obtained by including individual operational variables. Their small incremental contribution does not necessarily imply that they are physically unimportant; it may instead reflect information already represented by power demand and the relationships imposed by the controller. Discussing this distinction, including its implications for applications involving different control strategies, would help readers interpret the input-importance results.
In Section 4.3, line 846, the authors state that engineering flow models can provide local inflow and wake conditions, and therefore mean power, but do not contain turbine operational variables. It is unclear whether “mean power” refers to power demand or predicted electrical power. Please clarify this terminology and its relationship to the power quantities used as surrogate inputs. In particular, it would be helpful to explain under what conditions an inflow-based mean-power estimate could substitute for a required input.
3. The differences between the input distributions in Figure 6 should be considered when interpreting the validation and testing errors.
The testing dataset appears to have a higher minimum wind speed and a lower maximum value of the turbulence-related quantity than the training and validation datasets, indicating different coverage of the distribution tails. Although independent testing is a strength of the study, differences in aggregate NRMSE between validation and testing may reflect both model generalisation and dataset composition.
It would be useful to discuss how these distributional differences may contribute to the reported errors, which may be difficult to be distinguished from aggregate metrics alone. If readily available, a comparison over common wind-speed and turbulence ranges, or a limited breakdown of errors by operating condition, could clarify the interpretation.
The discussion could also explain what the testing results establish about performance near the limits of the investigated input range and what remains untested for sparsely represented conditions or input combinations. This would help users understand the additional validation needed before applying the framework under conditions that differ substantially from those represented in the current databases.
Citation: https://doi.org/10.5194/wes-2026-191-RC3
Data sets
Dataset and code supplement to the scientific paper "Design and implementation of a location-agnostic and control-oriented load surrogate model for wind turbines" Adrien Guilloré, Andreas Vad, Filippo Campagnolo, Carlo L. Bottasso https://doi.org/10.5281/zenodo.23372270
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 476 | 164 | 30 | 670 | 35 | 25 |
- HTML: 476
- PDF: 164
- XML: 30
- Total: 670
- BibTeX: 35
- EndNote: 25
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This manuscript presents an extensive study of reduced-order surrogate models for predicting wind turbine fatigue loads. Although the analysis is detailed, I have a fundamental concern about the generality attributed to the results. The training, validation, and nominally independent testing databases are all generated numerically using essentially the same modeling framework. The training database uses three IEA 3.4 MW reference turbines simulated with OpenFAST/FAST.Farm, and the independent test case changes this to an irregular five-turbine layout but retains the same turbine, aero-servo-elastic model, controller framework, FAST.Farm wake model, and simulation assumptions. Thus, the test set is independent in terms of wind farm layout, but it is not independent in terms of the underlying turbine physics or load-generation model.
This distinction is important because the manuscript makes broad claims concerning a “general” and “location-agnostic” load surrogate and derives specific recommendations concerning the necessary inflow representation, number of turbulent realizations, and required training database size. The study does demonstrate that these choices work consistently within the numerical environment that generated both the training and test data. It does not establish that the same conclusions hold for a different aeroelastic model, turbine platform, controller, wake representation, or, most importantly, measured turbine loads.
Indeed, the authors acknowledge that the method remains turbine- and controller-specific and that substantially broader validation using higher-fidelity simulations, wind-tunnel experiments, and full-scale field measurements is still required to establish robustness and transferability under realistic operating conditions. This qualification is difficult to reconcile with the much broader claims of robust, general load prediction and the prescriptive design recommendations presented elsewhere in the manuscript. The issue is not the amount of simulation data. The database is extensive. The problem is that a surrogate trained and tested against different realizations of the same underlying numerical model can principally demonstrate its ability to reproduce that model. Changing the wind farm layout does not provide independent validation of the fidelity of the predicted fatigue loads.
In my view, the principal claims of generality and robustness require independent physical validation that is not provided here. Obtaining that evidence would constitute a substantial new study rather than a revision of the present manuscript.