Downscaling of surface wind forecasts using convolutional neural networks

Dupuy, Florian; Durand, Pierre; Hedde, Thierry

doi:https://doi.org/10.5194/npg-30-553-2023

Articles | Volume 30, issue 4

https://doi.org/10.5194/npg-30-553-2023

© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/npg-30-553-2023

© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 30, issue 4

Research article

|

29 Nov 2023

Research article |

| 29 Nov 2023

Downscaling of surface wind forecasts using convolutional neural networks

Florian Dupuy, Pierre Durand, and Thierry Hedde

Download

Final revised paper (published on 29 Nov 2023)
Preprint (discussion started on 22 Jun 2023)

Interactive discussion

Status: closed

RC1:
'Comment on npg-2023-13', Anonymous Referee #1, 05 Jul 2023
Summary

The authors build on downscaling methods using convolutional neural networks for wind forecasts over south-eastern France. Here, they focus on exploring different output variables and loss functions. They find that there is not one that is better than the others in terms of both direction and speed simultaneously. But combining two different CNNs one for direction and one for speed results in a better performance. This is then analyzed using error metrics in a quantitative and qualitative way.

The manuscript is well written and gives new insights trough the extensive evaluation. Parts of the method section need to be improved to better understand which variables are used and which data is used for training and testing. Finally, the results that the authors already have for unseen areas should be included in the study.

Comments

In the abstract, could you use % to explain how the MAE is reduced instead of just numbers?

In section „2 Methods“ could you state your objective and which variables are used to predict the speed and direction in a small paragraph?

In section „2.2 Training data“ could you state which data you used for training and which for testing?

L 119: 32x32 data and 288x288 data, by data do you mean grid points?

L 119: which domain corresponds to the HR grid of 288x288 and why do you need this larger domain which is than cropped?

Figure 3: where are N_I and N_O (the number of input variables and the number of target variable) specified?

L 131: Please explain what you mean by „speed“ and „direction“

L 138: „We tested their approach“. Please rephrase such that it is clear that you trained another CNN using the described loss.

L 150: which other way is used to compute the wind speed?

L 160 onward: Why do you incorporate the condition u^2+v^2=1 in the loss and not predict u and compute v or have it built in the neural network architecture?

In section „2.4 Wind forecast evaluation“ could you state all the formulas for the evaluation metrics used?

In section „2.5 Computational considerations“ could you state how long you trained the neural networks and on which GPU?

In section „3.3 Wind climatology at specific sites“, over which period is the climatology computed? And why only at 2 different locations?

Figure 11: Please specify „whole period“.

The last sentence in the conclusion states that you have more results for evaluating the CNNs over unseen areas. If you already have these results why not include them?
Citation: https://doi.org/10.5194/npg-2023-13-RC1
- AC1: 'Reply on RC1', Florian Dupuy, 21 Aug 2023
  
  Please see the Author Comment in the attached pdf.
  
  Citation: https://doi.org/10.5194/npg-2023-13-AC1
RC2:
'Comment on npg-2023-13', Anonymous Referee #2, 20 Jul 2023

1. General comments

Dupuy et al. explored methods for grid-to-grid downscaling of surface wind. Building on the latest contributions in the field, which are well documented throughout the manuscript, they evaluate a variety of approaches. In doing so, they have also included key components of previous works. This contributes to a certain continuity in the literature, which is appreciated and helps focus new efforts. Rather than proposing a new architecture, the authors focus on the different modeling choices in terms of target variables and loss functions. They find that specific approaches perform best for either wind speed or wind direction, but not both at the same time. They show how the approaches can be combined to yield the best results on their evaluation.

The manuscript is well-written and easy to follow. The methodology is solid, although some minor aspects could be improved. Overall, a very valuable contribution to the field. I accepted with minor revisions, see section 2.2.

2. Specific comments

2.1 Discussions/clarifications

- Predictors: in the final results have you used time-related variables, such as cosine and sine components of the hour of the day? If not, have you considered them during your study? In combination with topographical predictors, they might help model the diurnal cycle.

- Figure 11a and b: I might be wrong, but 11b (CNN) seems to have a chessboard-like pattern, however 11a (WRF LR) does not. Since it is my understanding that CNN results use WRF LR as input, I find it a bit strange. Are you sure that the same interpolation method (bicubic) is used in both cases?

- u^2 + v^2 = 1: could you not strictly enforce this condition directly in the architecture, by having the model predict "u(x)" and "sign(x)" in "v = sign(x) sqrt(1 - u^2)"? If possible, it could be a nice addition to the manuscript.

- Area of study: given the focus on complex terrain, I question whether the D3 domain is the best choice. Moving it a bit further to the east would have included a more diverse set of topographical features (although would have excluded Mont Ventoux, ironically), possibly giving more insights and highlighting the benefits of the presented methods even more. This does not change the value of the manuscript, but it might be something to consider for future contributions.

2.2 Minor revisions requested:

- 288x288 domain: this is not the same as the D2 domain, correct? I would make this more clear, and maybe include this domain in Figure 1.

- Please provide more details on the cross-validation strategy.

- Figure 4: it would help to have the label of the best-performing model in boldface.

- Overall wind speed climatology: while I appreciate the in-depth analysis at the two specific sites, and understand its value, particularly for the qualitative evaluation, I believe more domain-level (or at many randomly selected points, if the size of the data is a constraint) quantitative analysis is needed to complement the verification metrics. For instance, it would be nice to see scatter plots or conditional quantile plots (see Wilks 2011) for wind speed for the entire spatial and temporal domain.

- The sub-optimal generalization capability outside the D3 domain is to be expected for this methodology. If you already have some relevant results, they should be included and discussed. In the future, it will be interesting to see how domain-agnostic models compete with domain-specific models.

3. Technical comments

- L62-65: a bit convoluted. I suggest rephrasing it as "31 km to 9 km (ratio close to 3) in Höhlein et al. ((2020)" etc.

- L326: "The diurnal cycle remains" Is this referring to the "cycle" in the MAE? I find it a bit confusing. Please rephrase.

Citation: https://doi.org/10.5194/npg-2023-13-RC2
- AC2: 'Reply on RC2', Florian Dupuy, 21 Aug 2023
  
  Please see the Author Comment in the attached pdf.
  
  Citation: https://doi.org/10.5194/npg-2023-13-AC2

Peer review completion

AR: Author's response | RR: Referee report | ED: Editor decision | EF: Editorial file upload

AR by Florian Dupuy on behalf of the Authors (21 Aug 2023) Author's response Author's tracked changes Manuscript

ED: Publish as is (08 Sep 2023) by Stéphane Vannitsem

AR by Florian Dupuy on behalf of the Authors (19 Sep 2023)

Short summary

Forecasting near-surface winds over complex terrain requires high-resolution numerical weather prediction models, which drastically increase the duration of simulations and hinder them in running on a routine basis. A faster alternative is statistical downscaling. We explore different ways of calculating near-surface wind speed and direction using artificial intelligence algorithms based on various convolutional neural networks in order to find the best approach for wind downscaling.