Flexible Load Multi-Step Forecasting Method Based on Non-Intrusive Load Decomposition

Round 1

Reviewer 1 Report

This paper proposes a multi-step flexible load prediction model based on a non-intrusive load decomposition technique and an Informer algorithm. The non-intrusive load decomposition technique and deep learning method are used to extract and map features from the flexible load to the total load, and the total load is decomposed into the flexible load and the residual load by the CNN-17 BiLSTM model. Subsequently, the Informer model is used to predict the flexible load and the residual load in multiple steps, respectively, and the prediction results are summed to obtain the overall prediction results.

 

From the abstract, we can not find the contributions of the paper, such as the precious results, only the improvement but how much improvement, and from the conclusion, the paper summary did not find. In addition, the data collection and format should be described in the paper. Authors are suggested to define ‘non-intrusive’ first; in addition, some similar papers are not cited, such as “Non-intrusive load decomposition based on CNN–LSTM hybrid deep learning model”  

1. On page 7, formula 14 is MAPE, not MAE

2. On page 7, the m value of Formula 15 needs to be outside the root square.

3. Conclusion is used to summarize the method of the paper and indicate the accuracy of your methodology. The advantages are better to move to section one. Authors need to show their future works in this section.

4. There are two sections 2 in the paper. One is on page 2, and the other is on page 5.

 

5. On page 8, the section title of 3.3 is Algorithm 1. However, that is not an algorithm. A series of steps to do the experiments.

Please carefully check the writing such as many definite articles are missing.

Some sentences are too long to read.

Author Response

Thank you for your comments. I have answered your questions in the document and corrected them in the appropriate places in the article.

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors propose a multi-step flexible load prediction model based on a non-intrusive load decomposition technique and Informer algorithm. The issue of load forecasting methods is an interesting and very important topic. However, in the introduction, there is a lack of directly described methods for the results of which the authors will compare the obtained results in their study. From the mathematical point of view, the description of the methods used is very vague, no consistent notation is used.

There are many small inaccuracies such as those listed below.

Fig.2, in the upper part instead of x_t should be x_{t-1} (the input to the hidden layer at a given time)

Fig.3, in the last column instead of h_t should be h_{t+1}

Formulas (8-10), what are L and i? In (9) I suppose it should be 2j-1 instead of 2j+1 (typically the range is 1 to d, not 2 to d+1). There is a lack of range of p in the sum in (10). Why PE in (8) and (9) has two indices, i.e. (pos,2j) or (pos,2j+1), and in (10) only one (L_{x(t-1)+i})?

Formulas (11-12), what is d_k or AB? The expressions are taken out of context (incomprehensible).

The results of the experiments presented in the figures are hardly legible.

 

Punctuation (especially in sentences that contain mathematical formulas) and captions under figures and tables should be corrected.

Author Response

Thank you for your comments. I have answered your questions in the document and corrected them in the appropriate places in the article.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors have corrected their errors and enhanced their content. I suggest it is acceptable.

Author Response

Thank you very much for your approval.

Good luck to you.

Reviewer 2 Report

I don't understand the two sentences (in the Abstract, lines 19-21) "The prediction coefficient of 0.9813 is 0.0069 when the number of prediction steps is 1, and 0.067 when the number of prediction steps is 20. The prediction coefficient of 0.9646 for the flexible load air conditioner when the model is applied to the dataset 2". Similar sentences are in lines 109-111.

The description of mathematical formulas has been slightly improved, but there are still some unexplained details, especially in formulas (10) and (11). The description " Q, K, V are three matrices of the same size obtained by linear transformation of the input eigenvariables" is insufficient. I suggest adding a reference to the literature where these formulas come from.

The sentence (lines 216-219) " After several operations of the multi-head probability sparse self-attentive module and the "distillation" mechanism module, the encoder can output one intermediate result, which can reduce the length of each sequence by half, saving the memory overhead of the encoder and improving the computation speed." is hardly intelligible.

Formula (15) RMSE is incorrect, RMSE should be the square root of the MSE!

Why mean absolute error (MAE) is multiplied by 100%? I wonder if, in Tab.3 or Tab.4, the values of MAE are described as percentages.

When authors write about "REDD dataset house6" (line 266), it should be given some explanation or link to the database.

The results presented in the figures have been slightly improved, although the different parts of the graph curves still are not fully recognizable.

In this revised version, the authors added to the paper new sentences that are incomprehensible and contain grammatical and logical errors.

Author Response

Thank you very much for your comments. I have made changes based on your comments.

Author Response File: Author Response.pdf