Detecting Nontechnical Losses in Smart Meters Using a MLP-GRU Deep Model and Augmenting Data via Theft Attacks

Round 1

Reviewer 1 Report

This study provides an efficient deep learning method for detecting NTL in electricity losses. Overall, this paper is well written. Sufficient references to previous studies and related work are provided. In addition, the authors provide a detailed explanation of the method they used. I recommend the authors make the following changes.

 

1. I believe that the current title of the paper does not show the main innovation of this study. The biggest innovation in this work is the data augmentation part, where the authors created synthetic theft attack data to balance the training dataset, which no one seems to have done before. Considering that both MLP and GRU are very old models, the hybrid MLP-GRU model is not a significant innovation at all.

 

2. The abstract of the paper tries to include too many trivial details. For example, the randomized search of hyperparameters is very common in machine learning and there is no need to emphasize it in the abstract.

 

3.  Instead of focusing on comparing MLP-GRU with MLP-CNN and MLP-LSTM, the authors should demonstrate the necessity of combining MLP and GRU. For example, what happens if only GRU is used for the auxiliary data and EC data? Or, what happens if no auxiliary data & MLP is used? In other words, the author should convince the reader that it is very necessary to use both parts of the model and both types of data.

 

4. The authors should be very clear about how many training and test data examples they have. If the dataset is not large enough, there may be a "data scarcity" problem. In this case, the authors should compare the proposed model with traditional machine learning models rather than deep learning models that requires more parameters.

 

5. I suggest that the authors combine the introduction and related work sections. I realize that this may be the standard way of writing conference papers, but for journal papers, such a long background introduction seems redundant.

 

6. Also, it is not necessary to mention every previous study and criticize each. Many previous studies use traditional machine learning models, while this paper focuses on deep learning - they face a different size of training set.

 

7. Minor issues:

 

a) Line 448. If my understanding of the paper is correct, it should be that the MLP receives the auxiliary data and the GRU receives the measurement data.

 

b) Line 39." While Paksitan faces .... of NTLs". This is an incomplete sentence. Consider changing it to ", while ......"

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This study aims to perform NTL detection using smart meters' data with a deep learning model. Listed below are some points that I would like corrected.

 

 

1) In lines 36 and 37, please provide references.

2)In line 132, Give a numerical value for the low detection rate you mentioned.

3) Correct the typo on line 179 (performancedegrades).

4) In line 167, Give a numerical value for the low detection rate you mentioned.

5) In line 179 pleasae provide references.

6) Between lines 166-179, write the deficiencies of the studies you mentioned in the literature by explaining them in a demonstrative way, not with general comments.

7) This applies to most references in general. For example, you say it provides low detection rates. Write down this low detection rate you mentioned, giving it a numerical value. What should be the expected value according to the literature? Is there a scale defined as low detection and high detection rates in the literature? Detail these parts.

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx