the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 01 Jun 2026
Accelerating regional wind energy assessment with deep-learning surrogates of WRF wind farm simulations
Abstract. Downwind wake effects are becoming increasingly important as wind farms grow larger and turbine capacities increase. Mesoscale weather models with wind farm parameterizations have emerged as a key tool in modelling long-distance wakes between wind farms. However, they can be too computationally expensive for evaluating dozens or more layouts considered in regional planning and optimization. In this study, we develop deep-learning surrogate models that reproduce the power losses and wind speed deficits caused by turbine layouts in mesoscale simulations at a fraction of the computational cost. The models combine atmospheric inputs from free-stream Weather Research and Forecasting (WRF) model simulations with turbine layouts to predict spatial power fields produced by WRF when the wind farm parameterization is activated. First, convolutional neural networks (U-Net) are developed as deterministic surrogates and achieve strong accuracy on two unseen scenarios. Second, diffusion-based models are developed to generate predictive ensembles and quantify uncertainty, including a residual diffusion model that learns the error of a deterministic U-Net prediction. Overall, the all models show a strong ability to predict wind power, both on a per-grid cell basis and aggregated across wind farms. The U-Net model strength shows sensitivity to the predictand (capacity factor vs. normalized power output), the combination of predictors (wind speed, wind direction, turbulence, and temperature), the number of training scenarios, and the type of loss function. Among the probabilistic models, DDPM provides the best calibrated ensembles, whereas residual diffusion yields more accurate point predictions and better farm-level bias control. These results demonstrate that deep-learning surrogates can enable rapid and cost-effective evaluation of candidate wind farm layouts, while also supporting uncertainty-aware planning-stage assessment.
- Preprint
(5388 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on wes-2026-87', Anonymous Referee #1, 21 Jun 2026
-
RC2: 'Comment on wes-2026-87', Anonymous Referee #2, 19 Sep 2026
General comments
This study investigates the use of deep learning models (a U-Net, and two diffusion model variants) to predict regional-scale wind turbine power, accounting for long-range wake effects. The deep learning models are trained using data derived from a mesoscale weather prediction model with a wind farm parameterisation. This addresses a significant problem in wind energy modelling, namely the computational cost of evaluating a large number of layout scenarios using mesoscale models. The study presents a novel approach, and is methodologically sound. I recommend that the paper will be suitable for publication after revisions focusing mostly on methodological clarifications and deeper discussion of the limitations of the current approach.
Specific comments
- The introduction should make clearer the distinction between farm-scale simulation (where engineering models and most existing deep learning wake modelling approaches are broadly applicable), and regional-scale simulations where mesoscale modelling is appropriate (and as noted by Pryor et al (2024), is perhaps the only appropriate approach). Here, you also refer to "layouts", which to me implies a farm-scale focus which actually under-sells the benefits of your approach, namely its applicability to long-range interactions.
- Missing or suggested literature:
- At lines 32-34, it would be helpful to refer to large-scale studies which have encountered this computational cost problem and which would most benefit from a surrogate model approach; there are numerous examples from Europe e.g. Baas et al (2024), Dörenkämper et al (2025), Akhtar et al (2021)
- At lines 75-77, the reasons for contrasting with downscaling problems are not clear. If this is to draw on the literature on deep learning-based downscaling as it relates to the methods you use in this study, this should be expanded.
- I suggest consulting the new review by Dong et al (2027) for further recent literature on similar problems, in particular refs 98-104 therein, some of which are closely aligned with your study. The novel contributions of your work should be expressed in this context.
- Methods and data preparation:
- Justification should be given for the chosen WRF configuration, e.g. by reference to validation studies. Why is 1km resolution used, and what is the distribution of the number of turbines per grid cell? This is important to understand the limitations of the power modelling approach.
- More information is needed about the types/models of turbines simulated. Are all turbines the same? What is the range of hub heights and rated powers? As much detail as possible should be provided, within the data confidentiality limitations. Turbine capacity range is briefly mentioned in the discussion but should be made clear earlier; it is a key methodological assumption that all turbines can be treated equivalently.
- It should be made clear how the TMY was broken down for the WRF simulations. You mention 3-day simulations in the context of computational cost; is the training also broken down in this way, and how did you arrive at this duration? Is there a spin-up period?
- The authors should justify, and discuss the implications of, using only a single instantaneous snapshot as input to the surrogate model
- The authors should justify the choice to use ensembles of 10 members for the diffusion models
- Sensitivity analysis:
- It is not sufficiently clear in section 3 (and table 2) what your baseline set of inputs are. E.g. you say "Beyond wind, we consider hub-height air density and turbulent kinetic energy, as well as the potential temperature gradient between 200 m and 20 m." It is not clear that these are excluded from your default model. Lines 354-355 then imply that all variables are in fact included for the "main model comparisons"; please clarify.
- Discussion around fig. 4 is slightly difficult to follow, and seems almost contradictory. Your discussion suggests the improvements are small, "indicating only limited incremental benefit", and at line 350 onwards you give further reasons why these improvements shouldn't necessarily be trusted. But you then conclude that all inputs should be retained anyway. Please clarify this discussion. Also, since you decide to include all inputs anyway, a more conventional way to perform this investigation would be to start with all inputs, and perform an ablation study to investigate the influence of each. You should justify your investigation approach.
- It would be informative to experiment with reducing the temporal range of the training data, in comparison to varying the number of layout scenarios seen in training. Given your fig. 3 results, and that the number of scenarios is several orders of magnitude smaller than the number of time snapshots, it seems likely that a better use of computational resources would be to decrease temporal coverage and increase layout diversity; this would be a useful insight to provide, if you find it to be true.
- Results are mostly assessed in terms of power, rather than wake-induced losses. I suggest that results should either be expressed in terms of wake-induced losses, or compared against a baseline no-wake model. At the least, the authors should explain why the error metrics are not exaggerating the model performance.
- Fig 8 and surrounding discussion: this in particular would likely be more informative if investigated as wake-induced losses (or "normalized wind power" as in section 4.2.6), rather than power. It is easy to get close to the "truth" curve armed only with a WRF simulation (no farms) and the turbine power curve. The wake-induced losses are the important output, and would show the differences between your modelling approaches much more clearly when expressed as a timeseries.
- Discussion:
- The authors are correct to identify lack of domain generalisability as a key limitation. This should be made clearer near the beginning of the discussion, since this undermines the "removing a key cost barrier" claim. Given the workflow as presented in this study, the method relies on a large volume of training data being generated on the same domain as the intended application. Given the range of ongoing and future work discussed, I recommend re-framing the current study as a preliminary or proof of concept towards the stated aim of removing the computational cost barrier. The similarity between the training and evaluation scenarios should also be made clearer throughout the results and discussion, as this is a key caveat which limits the strength of the conclusions which can be drawn.
- Missing discussion points:
- Limitations of the underlying WRF-WFP model. E.g. where more than one turbine is present in a single cell, wake effects are not properly represented. It would also be of benefit to review WRF-WFP validation studies in the introduction to add credibility to the overall workflow.
- You mention in the results section that the models perform better under stable wind conditions. It seems likely that this is related to the use of a single instantaneous snapshot being delivered as input to the surrogate model. The limitations of this choice should be discussed.
- You mentioned earlier the contrasting computational cost associated with inference of the U-Net and diffusion models. This should be revisited here to inform your comparison of the benefits and weaknesses of the two approaches.
References
Pryor, S.C. and Barthelmie, R.J., 2024. Wind shadows impact planning of large offshore wind farms. Applied Energy, 359, p.122755.Dong, Z., Zeng, W., Deng, X. and Yang, K., 2027. Machine learning for wind farm wake modeling: Output-driven taxonomy, architectural inductive biases, and benchmarking. Renewable and Sustainable Energy Reviews, 243, p.117359.
Baas, P., 2024. Winds of the North Sea in 2050—Public Final Report.
Dörenkämper, M., Vollmer, L. & Lakdawala, Z., 2025. Simulating the impact of offshore wind expansion on yield and efficiency in the north sea by 2050 (https://www.bsh.de/DE/THEMEN/Offshore/Meeresfachplanung/_Anlagen/Downloads/IWES_Bericht.pdf)
Akhtar, N., Geyer, B., Rockel, B., Sommer, P.S. and Schrum, C., 2021. Accelerating deployment of offshore wind energy alter wind climate and reduce future power generation potentials, Sci. Rep., 11, 11826
Technical corrections- Abstract: "Overall, the all models show..."
- Line 480: "larger when atmospheric covariate changes rapidly"
Citation: https://doi.org/10.5194/wes-2026-87-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 425 | 157 | 86 | 668 | 97 | 78 |
- HTML: 425
- PDF: 157
- XML: 86
- Total: 668
- BibTeX: 97
- EndNote: 78
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This seems to be an important study with relevance to wind farm planning. So the reviewer started reading with enthusiasm but found himself stuck after a short while.
The Introduction gives a short abstract of existing research citing a few overview papers. The open caveat is analysed and the desired solution is named: a “U-Net” and a “diffusion-based probabilistic surrogate model”. Unfortunately, no definition of a U-Net or a diffusion model is given. As these terms seem to define the core of the solution, some sentences on U-Nets and diffusion models should be in order.
Due to the missing definition of these two terms the reviewer has not understood the entire second paragraph of Subsection 1.1 (lines 78-90). Section 1 needs reformulation in order to provide access to the content of this manuscript also for non-specialists.
The same applies to Section 2 which is totally written in a formal language. Although this might provide a precise notation of the method it is inaccessible for non-specialists. The link to reality seems to be the term “free-stream atmospheric covariates”, another term which is not explained (lines 214-220 and Table 2 offer something, but quite late).
Nothing is said in the introduction whether the authors refer to onshore or offshore wind farms. It is only from the first sentence in Section 3 that the reader learns that the study refers to onshore conditions. It would have been wise to address this aspect much earlier, because wakes are much more intense and larger in offshore environments than in onshore ones. The authors should amend the introduction respectively.
By the way, the method should be applied to the seas around the British Isles or to the German Bight. Here, losses due to wakes are much more relevant than the onshore example given in this study. This would be the true proof of the concept presented.
Figure 1 should have axes in physical dimensions (kilometres). Otherwise, the given resolution of 1x1 square kilometre of the WRF model cannot be compared to the size of the wind farms depicted in Figure 1. Appendix B can be skipped, as it does not provide any meaningful additional information,
I stop here making any further detailed comments on this manuscript. The manuscript turns out to be a very technical one and is, unfortunately, not digestible for non-specialists (the reviewer is a meteorologist with special interest in wind energy for decades). Only lines 492-520 in the Conclusions give a little taste how the presented work could refer to the real world of a wind engineer or an energy meteorologist.
Although I think that I understand the importance of the presented research, I can only recommend the publication of this manuscript if the Introduction is extended in order to inform the readers in a more detailed way on the purpose of the study and the linkage between this formal research and the reality. This should start, e.g., by using the term “meteorological variables” instead of “atmospheric covariates”. Maybe, a wind engineer or a meteorologist could be added to the team of authors.