Enhancing Wind Turbine Power Forecast via Convolutional Neural Network

Round 1

Reviewer 1 Report

This paper proposes a deep learning method,  based on 2D Convolutional Neural Network (CNN),  for wind turbine power forecast.  The major novelty in this work is the application of 2D CNN via Gramian Wind field (GWF) matrix to time series data, and therefore is a welcome and fresh addition to the Electronics journal.   Here are a few suggestion to make the paper better,  that authors should address point by point.

 

1. The proposed method is similar to TCN (temporal convolutional network), which is also applying 1D CNN to time series. The authors should site TCN, and optionally applied  in this paper and compare its performance with the other DL networks , it is available for Keras in:

https://github.com/philipperemy/keras-tcn

An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling arXiv preprint arXiv: 1803.01271

And cite an important new engineering application there in deep learning in computational plasticity and viscoplasticity: Deep learning for plasticity and thermo-viscoplasticity, International Journal of Plasticity (2021), 136, 102852

2. Training is the most time consuming part of DL, and the authors should provide some computational training performance comparison too between their CNN based network and GRU and LSTM networks, preferably on GPU-s, but if GPU-s are not accessible then on CPUs.

 

3. Finally, not everyone is familiar with machine/deep learning and some brief intro how it works should be provided, if not in the manuscript, then in an appendix.

Author Response

Thank you so much for your time and suggestions. We have revised the manuscript according to your kind advice.   Comment 1: The proposed method is similar to TCN (temporal convolutional network), which is also applying 1D CNN to time series. The authors should cite TCN, and optionally applied in this paper and compare its performance with the other DL networks , it is available for Keras in: https://github.com/philipperemy/keras-tcn   An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling arXiv preprint arXiv: 1803.01271   And cite an important new engineering application there in deep learning in computational plasticity and viscoplasticity: Deep learning for plasticity and thermo-viscoplasticity, International Journal of Plasticity (2021), 136, 102852.   Response: Thanks for this kind suggestion. We have applied TCN and compare its performance in Line 1-9, Page 10. We also cited the above article in References[24].   Comment 2: Training is the most time consuming part of DL, and the authors should provide some computational training performance comparison too between their CNN based network and GRU and LSTM networks, preferably on GPU-s, but if GPU-s are not accessible then on CPUs.   Response: Thanks for the good suggestion. We have compared the different training performances on CPU in Line 10, Page 10.   Comment 3: Finally, not everyone is familiar with machine/deep learning and some brief intro how it works should be provided, if not in the manuscript, then in an appendix.   Response: Thanks for the reviewer's kind suggestion. We have explained how deep learning works briefly in Line 2-8, Page 2.

Reviewer 2 Report

The authors propose a machine learning based to enhance the wind turbine power forecast. In particular, what is interesting is that the authors encode their problem into image data and then adopt a convolutional neural network.

All in all, the article is interesting and deserves to be accepted. I have only the following considerations:

• can the authors provide the dataset adopted and the results obtained in a public repository like, for instance, Github?
• if I am not wrong, in lines 156,157 is defined the variable WS that does not look to be used elsewhere in the article

Author Response

Thank you so much for your time and suggestions.

Comment 1: can the authors provide the dataset adopted and the results obtained in a public repository like, for instance, Github?

Response: Thanks for this kind suggestion. We are regrettable that because of funding projects, part of the data and code is confidential in this paper. However, other data and code can be available from the corresponding author upon reasonable request.   Comment 2: if I am not wrong, in lines 156,157 is defined the variable WS that does not look to be used elsewhere in the article.   Response: Thanks for your kind advice. We have revised the manuscript in Line 7, Page 4. We deleted the variable WS and redefined the variable WE.