Deep-Learning Model Selection and Parameter Estimation from a Wind Power Farm in Taiwan^†

Round 1

Reviewer 1 Report

In this article, the authors focused on the determination of proper hyperparameters of DLN model using a Q-learning scheme for four models, TCN, LSTM, RNN, and GRU. The article is interesting and meaningful, and it has a clear aim. Some comments are given:

1- The English language must be double-checked.

2- Table 5 is a picture, which is very blurry. Please change it to a table like other tables. Also, its caption should be above the table not below it.

3- Figures 11-17 are blurry, please change them with high resolution.

Author Response

Q1) - The English language must be double-checked.

Response: The English writing of manuscript revision has been sent it to a native English speaker for editing service and grammar modification in my work. Minor typing mistakes and grammatical errors have been corrected in the revised article. (in brown color)

Q2) - Table 5 is a picture, which is very blurry. Please change it to a table like other tables. Also, its caption should be above the table not below it.

Response: Thanks for reminding us. Corrected it.

Q3) - Figures 11-17 are blurry, please change them with high resolution.

Response: Thanks. Corrected them with high resolution picture.

Author Response File: Author Response.docx

Reviewer 2 Report

The authors propose the development of deep-learning model selection with parameter estimation for predicting the power produced by a wind farm in Taiwan.

  

The paper is must be carefully rewritten and organized in some parts. The overall structure is solid, but I believe some comments are needed to improve the readability.

 

The authors should consider inserting a nomenclature table to make the reading easier and remove the notation/figures not strictly linked to the proposed method since the paper is already dense with information.

 

A list of the main comments/doubts/request is as follows:

 

1. Introduction

 

- line 70: typo in refs: [8-8]

 

- line 75-80: put them as a text, not as bullet points and also rephrase lines 83 and 88 accordingly 

 

- lines 85-86: the sentence is truncated

 

2. Relate Work 

 

In this section, there is enormous confusion with nomenclature. 

In addition, I would personally remove figures 2 and 3 and relative questions if they are not strictly correlated to the proposed methodology or serve to describe it appropriately.

 

- line 133 needs a reference 

 

-  Eqs (1-7): b_t, b_t, b_i, b_o, b_c are not defined

 

- line 163-165: c is defined twice with different names (candidate layer and memory state)

 

- Figure 2: what is \lambda? this figure needs consistency in notation with what was previously introduced

 

- Eq (9): the same symbols have been defined above (lines 162-166) with different definitions.

 

- Figure 3: there is a symbol partially covering the word Environment

 

- line 284: there is a link to Wikipedia

 

3. Determining the Hyperparameters of the Developed Model with Q-learning Scheme 

 

- line 333: there is a font change

 

- lines 449-450 should be put on a table or alone in the text 

 

4. Experimental Results 

 

- line 630 typo: “Figure 10 shows” : figure 10 is not the right one to show four figment prediction models

 

- How do you avoid the Overfitting problem?

 

5. Method Comparisons 

 

- Figure 17: what are the two plots? there is no distinction in the caption and the caption ends with a 0.

 

 - Table 11: bad formatting

 

Author Response

Q1) The authors should consider inserting a nomenclature table to make the reading easier and remove the notation/figures not strictly linked to the proposed method since the paper is already dense with information.

Response: 

1. Add a nomenclature table with a symbol list to represent math symbols for the proposed algorithm in front of Sec. 2.
2. Remove Figures 2 and 3 where the related information has appeared in the reference papers.

----------------------------------------------------------------------------------------

Q2) A list of the main comments/doubts/request is as follows:

Chapter 1.Introduction

• line 75-80: put them as a text, not as bullet points, and also rephrase lines 83 and 88 accordingly

Response: revised it to a text for lines 75-80. Also, rephrased lines 83 and 88 accordingly.

• lines 85-86: the sentence is truncated

Response: Correct the related sentences were truncated.

----------------------------------------------------------------------------------------

Q3) 2. Relate Work 

 In this section, there is enormous confusion with nomenclature. 

• In addition, I would personally remove figures 2 and 3 and relative questions if they are not strictly correlated to the proposed methodology or serve to describe it appropriately.

Response: Add a nomenclature table with a symbol list to represent math symbols. Accept the recommendation to remove figures 2 and 3 for clarity.

• line 133 needs a reference

Response: Added the related references [7-13] for sequence data processing DLNs to the text.

• Eqs (1-7): b_t, b_i, b_o, b_c are not defined

Response: added the following definition for four bias vector parameters in an LSTM unit.

 represents bias vector parameters to be learned during training in forget gate, input gate, output gate, and memory cell, respectively.

• line 163-165: c is defined twice with different names (candidate layer and memory state)

Response: Revised as represents the candidate (memory) cell state, c stands for the memory cell state. 

• Figure 2: what is \lambda? this figure needs consistency in notation with what was previously introduced

Response: l is the slope coefficient for the leaky ReLU function. For simplicity, replace ReLU(l) with ReLU in Figure 2.

• Eq (9): the same symbols have been defined above (lines 162-166) with different definitions.

Response: For consistency,    denotes the input vector to the neural unit through the whole text.

7) Figure 3: there is a symbol partially covering the word Environment

 Response: Adjusted it to no overlap in Figure 3

8) line 284: there is a link to Wikipedia

   Response: Deleted it.

----------------------------------------------------------------------------------------

Q4). Determining the Hyperparameters of the Developed Model with Q-learning Scheme 

• line 333: there is a font change

Response: Corrected it to Palatino Linotype throughout the whole paper.

2) lines 449-450 should be put on a table or alone in the text 

Response: Added Table 3 to adopt the basic features of wind turbines in the Chang-Bin industrial zone. 

----------------------------------------------------------------------------------------

Q5). Experimental Results 

• line 630 typo: “Figure 10 shows” : figure 10 is not the right one to show four figment prediction models

Response: Correct it to the right number of Figure, Figure 14.

 

• How do you avoid the overfitting problem?

Response: As shown in Figure 2, we added a dropout function in the dense layer to reduce model overfitting. In other words, the dropout function in the dense layer decreases the number of parameters to be learned and helps reduce overfitting.

----------------------------------------------------------------------------------------

Q6). Method Comparisons 

 1) Figure 17: what are the two plots? There is no distinction in the caption and the caption ends with a 0.

Response: Renumbering Figure 17 to Figure 15 due to removing Figures 2 and 3. Divided Figure 15 into Figure 15(a) and Figure 15(b). Figure 15 (b) is a partially enlarged image of Figure 15(a) used for discriminating the convergence error differences for four deep learning models. Delete the caption that ends with a ‘0’.

• Table 11: bad formatting

Response: Corrected it.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The authors have replied critically to all my questions/doubts.

I suggest a further minor revision to avoid any misleading sentences and typos.