Feature selection techniques for modelling tower fatigue loads of a wind turbine with neural networks

Movsessian, Artur; Schedat, Marcel; Faber, Torsten

doi:https://doi.org/10.5194/wes-6-539-2021

Articles | Volume 6, issue 2

https://doi.org/10.5194/wes-6-539-2021

© Author(s) 2021. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/wes-6-539-2021

© Author(s) 2021. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 6, issue 2

Research article

|

21 Apr 2021

Research article |

| 21 Apr 2021

Feature selection techniques for modelling tower fatigue loads of a wind turbine with neural networks

Artur Movsessian, Marcel Schedat, and Torsten Faber

Download

Final revised paper (published on 21 Apr 2021)
Preprint (discussion started on 02 Jan 2020)

Interactive discussion

Status: closed

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

- Printer-friendly version

- Supplement

RC1: 'Referee Comment 31/01/2020', Anonymous Referee #1, 31 Jan 2020
RC2: 'Review', Anonymous Referee #2, 10 Feb 2020
AC1: 'Response to referees # 1 and # 2', Marcel Schedat, 13 Apr 2020

Peer-review completion

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

AR by Marcel Schedat on behalf of the Authors (20 Apr 2020) Author's response Manuscript

ED: Referee Nomination & Report Request started (05 Jul 2020) by Gerard J.W. van Bussel

RR by Anonymous Referee #1 (23 Jul 2020)

RR by Anonymous Referee #2 (11 Aug 2020)

Suggestions for revision or reasons for rejection

Many thanks providing an updated version of the manuscript. I acknowledge that you have addressed many of the comments by the reviewers. However, further changes are required in my opinion.

General comments

1) As stated in the paper, the novel aspects of the paper, for example in relation to work carried out by Vera-Tudela et. al., is the application of NCA. However, this is not sufficiently reflected in the paper in my opinion. It seems to focus on reproduction of results from other studies instead. For example, the paper does not use the opportunity to compare strength and weakness of NCA against the other approaches applied. In addition, interpretation of results from NCA is missing while results from correlation analysis are discussed in detail for example. I addressed this discrepancy between justification of the paper and the actual content already in my last review, i.e. comment p.7, comment p.10, l.19-20 and general comment on literature survey which were not taken into account.

2) I also made a general comment regarding presentation quality. I still find that the language is imprecise at times and some of my comments in the last review addressed such issues. However, this was not taken sufficiently into account and some of my comments below still relate to imprecise language. In that respect my comments below are not comprehensive and I recommend to carefully revise the paper in this respect.

Specific comments:

page 20, line 30:
Rapid growth is not an explanation for a fierce competition in my opinion. Maybe it could be stated instead that there is still the need to reduce the cost of wind energy further and that improved monitoring solutions can potentially decrease O&M costs?

p.24,Table 1:
Thanks for your answer. Shouldn't the description of the dependent variables then read "Bending moment derived from gauge sensor located..." instead of "Gauge sensor located...", especially as the unit is given as kNm?

p. 24, line 10:
I suggest to write "This calculation..." instead of "This transformation..." because in the sentence before "transformation" refers to deriving bending moments from strains.

p. 26, line 6:
Is the reason to split the data set into partial and full load operation also that some of the applied methods (Regression, PCA) are linear methods, i.e. cannot handle non-linear relations? If so, this could be mentioned either here or in the section where the methods are explained briefly.

p. 28, line 20:
Thanks for your explanation. I agree that there is no general rule how to chose the network topology. Because of that, testing different network topologies is usually required important. I understand from your reply that this was done at least regarding the number of neurons in the single hidden layer and that can be mentioned in the text. If I misunderstood you reply, you should at least justify the choice of the network by the satisfying prediction accuracy that you experienced during testing. Currently this sections reads as if you did not care about the network topology which according to your reply was not the case.

p. 30, line 21:
In the paper the term 'feature' relates to an explanatory variable. I recommend to replace the word 'feature' by 'principal component' in this sentence to distinguish between these two.

p. 32, line 21:
Thanks for your explanation. I still think that the sentence is imprecise. What do yo mean by deviation? Do you mean a change of wind speed? Also the next sentence does not make sense.

p. 32, line 30:
Thanks for your reply. However, it does not address my comment. I was wondering why the pitch angle was selected by stepwise regression at all, as it does not correlate with the DEL. Same holds true for air density an wind direction. Please clarify.

p.34, line 3:
Do you refer to Figure 4 or Figure 5 here?

p.34, line 6:
A few lines above you already mentioned that normalized predicted vs measured DELs is shown in Figure 5.

p. 35, line 5:
It seems that the turbine is in operation also at wind speeds lower than 5 m/s. Units should be corrected from m/s^2 to m/s.

p. 35, line 9:
What to you mean by 'significantly higher results'?

Appendix 1-3:
Many thanks for this additional information. For a better overview and comparison of the methods there could be one table only, where a cell contains "0.9, b, c" if this variable was used by all approaches and "0.9, c" if it was used by correlation and NCA for example. In that way only one table is needed and it could maybe be integrated into the main body of the paper (not Annex).
Please also change 'still stand' to ' stand still' for consistency with the rest of the paper.

Hide

ED: Publish subject to minor revisions (review by editor) (30 Nov 2020) by Gerard J.W. van Bussel

AR by Marcel Schedat on behalf of the Authors (31 Jan 2021) Author's response Author's tracked changes Manuscript

ED: Publish as is (05 Feb 2021) by Gerard J.W. van Bussel

ED: Publish as is (11 Mar 2021) by Jakob Mann (Chief editor)

AR by Marcel Schedat on behalf of the Authors (17 Mar 2021)

Short summary

The assessment of the structural condition and technical lifetime extension of a wind turbine is challenging due to lack of information for the estimation of fatigue loads. This paper demonstrates the modelling of damage-equivalent loads of the fore–aft bending moments of a wind turbine tower, highlighting the advantage of using the neighbourhood component analysis. This feature selection technique is compared to correlation analysis, stepwise regression, and principal component analysis.