Cartographing dynamic stall with machine learning

Lennie, Matthew; Steenbuck, Johannes; Noack, Bernd R.; Paschereit, Christian Oliver

doi:https://doi.org/10.5194/wes-5-819-2020

Articles | Volume 5, issue 2

https://doi.org/10.5194/wes-5-819-2020

© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

Special issue:

Wind Energy Science Conference 2019

https://doi.org/10.5194/wes-5-819-2020

© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 5, issue 2

Research article

|

29 Jun 2020

Research article |

| 29 Jun 2020

Cartographing dynamic stall with machine learning

Matthew Lennie, Johannes Steenbuck, Bernd R. Noack, and Christian Oliver Paschereit

Download

Final revised paper (published on 29 Jun 2020)
Preprint (discussion started on 01 Jul 2019)

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

RC1: 'Review of 'Cartographing dynamic stall with machine learning'', Anonymous Referee #1, 04 Sep 2019
RC2: 'Review of Cartographing Dynamic Stall with Machine Learning', Tuhfe Gocmen, 06 Sep 2019
AC1: 'Author Responses to comments', Matthew Lennie, 04 Nov 2019

Peer-review completion

AR: Author's response | RR: Referee report | ED: Editor decision

AR by Matthew Lennie on behalf of the Authors (11 Nov 2019) Author's response Manuscript

ED: Referee Nomination & Report Request started (15 Dec 2019) by Katherine Dykes

RR by Anonymous Referee #3 (30 Dec 2019)

Suggestions for revision or reasons for rejection

The aim of this manuscript is to involve machine learning processes to identify time-dependent stall processes for airfoils undergoing dynamic motions. The study is relevant to the research community and publication is generally recommended, albeit with reservation for both minor and major corrections.

[Abstract] Why is airfoil stall bad for wind turbines when it limits the aerodynamic loads and therefore prevents structural failure?

[17] The authors are strongly encouraged to remove the URL to Google Books, since Google is not publisher. For example, the first reference to the book by Abbott & von Doenhoff was published in 1959, not 2012.

[20] The author's name is Strangfeld.

[33] This is an effective decambering of the airfoil.

[41] The spatio-temporal variation is in both span and chord.

[77] An airfoil can exhibit LE separation without a preceding TE separation (and light stall).

[79] Please use proper scientific language. "Spike" and "dump" are not appropriate English words to describe this process. "Spike" describes a time-instantaneous behavior, or impulse, whereas "dump" is a colloquial phrase describing an entirely different process than aerospace engineering.

[160] The authors do not mention which of the two experimental facilities dataset are subsequently used when and where later in the manuscript. My understanding is the the data provided by Müller-Wahl was made available before the machine learning process development was initiated. Why not solely rely on this dataset in the manuscript and save the towing tank data for a later publication?

[223] Figure 6 is never referenced in the text. Which dataset is graphically described in this figure? How is the silhouette (misspelled in the figure ordinate) score calculated? Is there a descriptive equation?

[Fig. 9] Does this figure describe the motions used from both experimental facilities?

[340] The authors declare that they are not going to "conduct a proper analysis" of available data, but instead save this for later. If the authors have no intention of publishing the results and conclusions in this manuscript, they should consider withdrawing the manuscript and submitting it elsewhere when the proper analysis have been conducted.

[343] In Figure 14, the legend obscures half of the data and should be placed preferably in the lower left corner.

A general comment is that the figures are in many instances placed far away from the discussing text. In some cases the discussing text (and the figure reference) is missing altogether. The captions of many figures are very brief and not appropriately descriptive. A reader should not have to find the text body to understand the content of a figure.

The manuscript appears to have been hastily prepared and not thoroughly proof-read for visual formatting. Furthermore, it uses two different datasets from two very different facilities (wind tunnel and towing tank), with two different test articles (thick airfoil vs. thin flat plate). The manuscript results and conclusions do not mention which data was used when and where. It is unclear to the reviewer why the two data sets were used and whether a comparison was intended, but omitted. The manuscript will be much better focused if one is omitted - this is strongly encouraged before publication.

Hide

RR by Anonymous Referee #1 (18 Feb 2020)

Suggestions for revision or reasons for rejection

Thank you for revising the article. I still have some comments:

#### Comment regarding training method

I have no doubt that the training approach you lay out in the conclusions can work in principle. But it is important to acknowledge that the empirical models will not be able to give any information in the regions where they fail. For example they do not give any information on deep stall dynamics relevant for vortex induced vibrations. So in this region, all of the training data will have to come from CFD or experiments. To use your metaphor, the street map provided by the empirical models is completely blank in these areas. Exactly because the empirical models fail here, these areas are the interesting areas to explore. But these are also the most complex cases where 2D aerodynamics are not enough and results will depend on the aeroelastic behaviour of the blade (including chord and twist distribution, airfoil families, prebend or sweep, deflection, structural properties), and on the operating conditions (in case of vortex induced vibrations these include angle of attack, inclination angle of the blade, wind shear, turbulence). The required number of very expensive blade resolved CFD training cases would be very large. It is likely that the CFD solver will have to be coupled to a structural model to predict lock-in where the vortex shedding frequencies get pulled towards structural frequencies, which increases the computational cost.
In conclusion a machine learning model can't be trained with cheap data exactly in the areas where the machine learning model would be most beneficial because the existing empirical models fail. Because you specifically touch on vortex induced vibrations in the conclusions, I think it is important to mention that the only relevant training data in that region is the high fidelity or experimental data and that lack of training data or high cost of training data can be an issue in the model training process. I think this would be a relevant thought to add to the conclusion.
The six training steps you detail in the conclusion seem to assume that a 2D model that needs to be trained for different airfoils is sufficient. Due to the importance of the parameters I mentioned above, it will not be enough for vortex induced vibrations.

#### Comment Regarding Airfoil stall is bad

Maybe I was a bit too brief in stating why I think this sentence should be changed. You write in your answer that stall reduces lift. Exactly that can be a huge advantage in all the design load cases where the angle of attack can not be controlled, for example because the blade pitch is stuck or the turbine is idling without grid connection and thus the yaw angle can't be changed. Especially in the stand still cases, the angle of attack can basically reach any value. These cases can be design driving. But if the lift coefficient would just increase with a constant slope for increasing AOA the extreme loads would certainly be much higher than if stall is present. So stall is necessary for the current wind turbines to survive a storm in stand still or failure of the pitch system.
I also don't really see the necessity for judging a phenomenon as bad or good. You could simply start the abstract with 'Once airfoil stall has set in...'. I think the first sentence is not necessary, doesn't add anything to the article and can easily be attacked. That is why I suggested to remove it.

##### Cycle to Cycle Variations
I can't see the benefit of using the term stability as something that would change cycle to cycle. If the operation would indeed change between stable and unstable in different cycles, on average the damping must be close to zero. There will likely be no single cycle that deviates far from zero and no significant vibrations building up within a cycle. The stability of an operating point (which would then be approximated as the average of the damping over many cycles if the system is non-periodic) on the other hand can be used to make a statement if vibrations are damped out or if vibrations are building up. This seems like an important information to me, much more important than if a small amount of energy is going into the system during a single cycle where the damping is slightly negative.

##### Probability distributions
I think it would be easier to compare the different figures if they were plotted using an x and y-axis.

##### Figure 9
Figure 9 is not referenced in the text. Is this Figure necessary? If yes then you should add a discussion of it in the text, if no it should be removed.

Hide

ED: Publish subject to technical corrections (20 Feb 2020) by Katherine Dykes

ED: Publish subject to technical corrections (20 Feb 2020) by Carlo L. Bottasso (Chief editor)

ED: Publish subject to minor revisions (review by editor) (23 Feb 2020) by Katherine Dykes

AR by Matthew Lennie on behalf of the Authors (04 Mar 2020) Author's response Manuscript

ED: Publish subject to minor revisions (review by editor) (10 Apr 2020) by Katherine Dykes

AR by Matthew Lennie on behalf of the Authors (22 Apr 2020) Author's response Manuscript

ED: Publish subject to technical corrections (03 May 2020) by Katherine Dykes

ED: Publish subject to technical corrections (05 May 2020) by Joachim Peinke (Chief editor)

AR by Matthew Lennie on behalf of the Authors (07 May 2020) Author's response Manuscript

Short summary

This study presents a marriage of unsteady aerodynamics and machine learning. When airfoils are subjected to high inflow angles, the flow no longer follows the surface and the flow is said to be separated. In this flow regime, the forces experienced by the airfoil are highly unsteady. This study uses a range of machine learning techniques to extract infomation from test data to help us understand the flow regime and makes recomendations on how to model it.