the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 10 Apr 2026
Synthetic generation of long turbulent wind time series using hindcast model forcing for offshore wind farm simulation
Abstract. Offshore wind energy is crucial for the transition to a low-carbon society, and accurate modeling of turbulent wind fields is essential for the design and operation of offshore wind farms. This study aims to bridge the gap between mesoscale and microscale wind fluctuations to generate long time series that are statistically and spectrally representative of real observations, capturing the non-stationary nature of turbulence. Mesoscale data from NORA3 is combined with microscale spectra from Cheynet et al. (2018) using methodologies from Veers (1988); Sørensen et al. (2002); Chabaud (2024a) and the splicing technique introduced in Chabaud (2024b). The validation process uses observational data from the FINO1 weather mast. The model accurately reproduces the wind statistics. The along wind turbulence intensity is within a 85 % confidence interval of ±0.02 for 2 h simulations. The model is performing slightly better in stable conditions. The spectral representation is also good for periods between 2 min and 24 h. There, a mesoscale term is added to the microscale model following Larsén et al. (2013) —fitted parameters are provided— to bridge the gap between the hourly resolution of NORA3 and the typical minute-scale microscale range. The good performances and low computational needs of the presented methodology open new possibilities for the modeling of turbulence intensity, for instance for forecasting.
- Preprint
(2795 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 08 May 2026)
- RC1: 'Comment on wes-2026-58', Anonymous Referee #1, 26 Apr 2026 reply
-
RC2: 'Comment on wes-2026-58', Anonymous Referee #2, 27 Apr 2026
reply
The comment was uploaded in the form of a supplement: https://wes.copernicus.org/preprints/wes-2026-58/wes-2026-58-RC2-supplement.pdf
-
RC3: 'Comment on wes-2026-58', Anonymous Referee #3, 28 Apr 2026
reply
The paper describes the stochastic generation of ABL turbulence time series at specified points from microscale spectra and NORA3 data using methods previously published by one of the authors (Chabaud 2024a and 2024b), the focus of this work is on the experimental validation of the methods using data from the FINO1 weather mast. The extension of the turbulence spectrum towards lower frequencies is very relevant and necessary for the prediction of fatigue loads using aero-elastic simulations.
Perhaps a general criticism of the applied methods from the perspective of the industrial application, which should not affect the decision of publication of this work: the described method does not preserve the spatial structure of an incompressible flow; it generates one-point time series, but no solenoidal 3D velocity fields (this compromise is also mentioned by the authors, e.g. . We as a wind turbine manufacturer have experimental evidence from instrumented turbines indicating that using divergence-free velocity fields model has a significant effect on certain fatigue loads, with the DEL predictions of the divergence-free turbulence (e.g. Mann 1994) being higher than that of stochastic spectral models without that property (Kaimal 1972, very similar spectra and identical TI otherwise). Similar observations can be made when the components of an initially divergence-free velocity field are scaled (resulting in div \vec u \neq 0) to model inhomogeneous turbulence. For wind turbine loads predictions, it would be very interesting to see if the e.g. the Mann 1994 or Syed&Mann 2024 models can be adjusted/constrained using NORA3 data. The authors do provide a very short discussion on that in Sect. 4.3 to support their choices made in the paper. Again, this is a general comment and outside the scope of this work, and should therefore not affect the decision for publication.
A comment on the writing style: there is a certain level of textual "bloat" (for instance, the description of the spectral splicing method is fully from Chabaud 2024b and covers almost two pages in this work) and a rather high dose of marketing vocabulary ('explore this "terra incognita"', 'The power of FLAggTurb', 'The absence of a clear spectral gap underscores the importance of the presented splicing methodology'). The latter might be due to the use of Generative AI for writing the paper, as stated by the authors. These are the main reasons for my evaluation of the presentation quality as "fair" only.
The description of the statistical and spectral validation methods in 2.3 are very good, and from what I can see the main contribution of this work over Chabaud 2024a&b. I agree with the choice of unnormalized spectra for comparing FINO1 time series to those predicted by the presented stochastic method, but perhaps the argument argument against using Cheynet 2018 normalization could be worked out a bit better: as far as I can see, Cheynet 2018 only discusses the relative normalization of the component spectra S_w/S_u, not the weighting of different parts of the spectrum (i-iv) relative to each other. I guess that is meant by "lacks meaning outside the microscale range".
In Section 3.1, I agree with the conclusions of the authors that the method reproduces the histogram of the longitudinal velocity component u (Fig 5A) and the friction velocity (Fig 5F). I do not fully agree with the "excellent agreement" for the standard deviations for the three velocity components, though. The authors correctly point out the "battlement" (jagged or peaky histogram). The authors attribute this to the binning in the friction velocity (estimated by equation (2) from NORA3 data) and atmospheric stability (estimated by equation (1) from NORA3). It would be nice to have more details on how the NORA3 data resolution affects the binning here. Sect 4.1 discusses this point a bit further (influence of refining the binning), but battlement also remains for the fine binning. Being on a "hill top" or in a "valley" due to some binning in the data upstream of the parameter estimation can significantly impact the usability of the proposed method for the estimation of turbine loads that are highly sensitive to the standard deviation of the velocity components.
It is interesting to see in Fig 6D that the vertical velocity fluctuations are underestimated vs. the two velocity components parallel to the ground. The vertical velocity component is usually computed from the 3D continuity equation, and corresponds to the deviation of the 2D divergence of the flow in a plane from zero. It seems that the intrinsic divergence of the flow field generated by the Veers method takes its toll here, otherwise the low-frequency fluctuations of u,v and w would be coupled.
The second part of Sect. 3.1 (MBE and RMSE) is very good. Besides for the wind speed bins < 5 m/s, the proposed method has a low bias and an acceptable level of RMSE, the latter about twice the limit for wind speed measurements defined by IEC 61400-50. That is very good for such a stochastic method, and would make it suitable for site-specific fatigue life time estimates. If possible, I would appreciate a discussion for the deviations at bins < 5 m/s; this velocity range is increasingly important for larger rotors due to the accumulated fatigue damage during idling.
The spectral parameters of the Cheynet 2018 model are obtained from a fit to the FINO1 measurements, the same test site is used to obtain time series for the validation of the model. The presented validation should therefore be seen as a best-case scenario, where the spectral parameters are known from the experiments and are used together with the NORA3 hindcast data. Additional spectral parameter estimates are needed for a3 and a4, which were obtained by fitting the Larsen 2016 model to the FINO1 data. In practice, using NORA3 data alone will require an estimation of these parameters at a specified location. Perhaps the authors can comment on either the generality of their parameter estimates, and/or on how these can be obtained from NORA3.
Sect 4.3 discusses the use of the Veers method vs. Syed & Mann, again using rather bold rhetoric. I would challenge the statement that lifting Taylor's frozen turbulence hypothesis is more important than discarding the divergence-free structure of a 3D incompressible flow. At least from the perspective of predicting wind turbine loads, the contrary is the case. The authors also use a 1m rotor size to validate their model against a point spectrum in FLAggTurb, removing the low-pass filtering of large rotors (Sect. 2.2.2). Notwithstandingly, the authors provide a valid argument about the difficult parameter estimation for the Syed&Mann 2024 method.
Overall, I would suggest publication conditional to a minor review. The authors recycle a large part of their methods section from their previous work (Chabaud 2024a & b), new is the validation against FINO1 time series and spectra. As some spectral parameters of the Cheynet 2018 model were obtained from FINO1 fits, the work represents to some degree a circular closure and the generality of the approach for an arbitrarily selected location in the NORA3 domain is left in doubt. This is also a practical limitation for the implementation of the proposed method in an industrial tool chain. I would suggest that the authors address this point in a revised manuscript, as well as the comments above. I furthermore expressed some disagreement with modelling choices and statements made by the authors, but I am certainly biased there as well and would not consider that as prohibitive for publication. If there is time for some textual improvements, I would strongly suggest a revision of the AI-generated text in some parts of the paper.
Citation: https://doi.org/10.5194/wes-2026-58-RC3
Data sets
Wind speed and direction, Atmospheric Stability, Turbulence Intensity, Turbulent Kinetic Energy and Sensible Heat Fluxes in the North Sea (FINO1) L. Pauchet et al. https://doi.org/10.60609/b9t9-0x55
Model code and software
FLaggTurb V. Chabaud et al. https://gitlab.sintef.no/ser-windfarmtools/flaggturb/-/releases/v2.47
FarmStream L. Pauchet and V. Chabaud https://gitlab.sintef.no/ser-windfarmtools/farmstream
Interactive computing environment
LouisPauchet/Long_Wind_Time_Series_Processing_Notebooks: 0.0 L. Pauchet https://zenodo.org/records/18621659
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 214 | 65 | 20 | 299 | 14 | 17 |
- HTML: 214
- PDF: 65
- XML: 20
- Total: 299
- BibTeX: 14
- EndNote: 17
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The authors present a novel methodology to incorporate long-term mesoscale fluctuations into the generation of synthetic turbulent time series for wind farm energy yield and load assessment. The proposed approach is validated against microscale measurements collected at the FINO1 offshore meteorological mast.
The reviewer finds the topic timely and relevant, addressing an important challenge in wind energy science. The proposed methodology is sound, well-structured, and consistently applied throughout the manuscript. The results are presented clearly, and the limitations of the approach are appropriately acknowledged and discussed.
However, the manuscript would benefit from a stronger emphasis on the novelty and advantages of the proposed method compared to existing approaches, particularly those related to mesoscale spectral modeling. This aspect should be more clearly articulated, especially in the Introduction.
Specific comments:
1. Abstract: The description of the methodology is overly technical and does not sufficiently highlight the practical modeling advantages or its improvements over the current state of the art. A clearer, more impact-oriented summary is recommended.
2. Section 2: Including a schematic diagram of the proposed workflow would greatly improve clarity. In particular, a visual representation showing how different datasets are combined, along with the key inputs and outputs, would help readers better understand the methodology.
3. Section 4: The methodology relies on the NORA3 dataset. It would be valuable to discuss whether the performance of the method is sensitive to the choice of hindcast model, and how results might vary if alternative datasets were used.
4. Language: The manuscript would benefit from careful proofreading. In particular, the Introduction contains several sentences with grammatical issues, including missing verbs and awkward phrasing, which affect readability.