Forecasting of Reactive Power Consumption with the Use of Artificial Neural Networks

Round 1

Reviewer 1 Report

The topic of the paper is forecasting of 3-phase unbalanced current consumption, which is very interesting topic.  Good news is that authors use a real dataset. In this case, the data set is related to an industrial building connected to a low-voltage network, located in Gliwice city in Poland.

The MATLAB with the Neural Network Toolbox was used to implement and test the neural network.

The paper should be improved. The paper is about time-series predictions, but references to open-source software machine learning methods, like extreme gradient boosting, k-nearest neighbor or random forest should be at least mentioned, see for example Tomčala, J. Towards Optimal Supercomputer Energy Consumption Forecasting Method. Mathematics 2021, 9, 2695. https://doi.org/10.3390/math9212695

and Julien Siebert, Janek Groß, Christof Schroth. A systematic review of Python packages for time series analysis. https://arxiv.org/abs/2104.07406

The popular open-source tools like Prophet or AtsPy (Automated Time Series Models in Python) should be mentioned. It is a pity that the paper does not include a link the analysed data set.

The following non-standard notation should be verified and fixed:

Pg. 5, Eq.3: The symbol PE of percentage prediction error can be interpreted as P*E, which is misleading. Thus, PE should be replaced to a simpler symbol, for example E.

The symbol of projected value “yp_i“ looks like y*p_i, which is misleading. Thus, “yp_i“ should be replaced to a simpler symbol, for example “p_i“.

 

Pg.6, line 186: The phrase “over-learning” is used in psychology. But in machine learning, people use “overfitting”. The final version of the paper should use phrases, which are used in machine learning.

The phrase “learning "by heart"” is not used in machine learning and should be deleted. Or authors should use a proper English reference as a proof.

Figure 8 uses different colors from yellow to blue. Authors should explain the meaning of the colors. Is it color connected to the prediction error? If yes, how?

About Figure 9 and the forecasting methodology: Authors should use the train/test/validation methodology: A different data set should be used for validation of the model, see:

https://stats.stackexchange.com/questions/346907/splitting-time-series-data-into-train-test-validation-sets

 

Author Response

Dear Reviewer

In attachment I'm sending answers to the questions, and at the same time I inform you that the article has been corrected.

Author Response File: Author Response.pdf

Reviewer 2 Report

Article 1675587 is devoted to solving the problem of predicting reactive power using artificial neural networks. For the calculation, a neural network of non-linear autoregression (NAR) was used for various sizes of the input vector and various numbers of neurons in the hidden layer to predict the generation of reactive power. Simulation results, compared with real measurements, confirm the possibility of predicting the reactive power course, useful for optimal planning of the reactive power compensation strategy.
In the current conditions, this topic is relevant both from a scientific and practical point of view. The material is presented in an interesting way.
There are no comments on the article.
May be published as is.

Comments for author File: Comments.pdf

Author Response

Dear Reviewer

Thank You for Your Review

Reviewer 3 Report

The authors focus their study on the problem of forecasting the reactive power consumption by adopting an artificial neural networks approach. Specifically, the authors adopt the nonlinear autoregressive neural network considering different input vector sizes and different number of neurons in the hidden layers in order to predict the reactive power generation.

 

The authors have provided a detailed analysis of the theoretical model that they adopt, as well as a very detailed set of numerical results in order to show the pure operation and the performance of the proposed framework.

 

The manuscript is overall well written and easy to follow and the authors  have well thought out their main contributions. The authors need to address the following comments in order to improve the quality of presentation of their manuscript, as well as they should consider the following suggestions provided by the reviewer in order to improve the scientific depth of their manuscript.

 

Initially, section one needs to be substantially rewritten in order to use more summative language to present the existing research works to the reader and clarify what is their main contribution, as well as what is the scientific research gap that the authors try to address.

 

In Section 2, the authors need to discuss several existing approaches from the controls point of view, such as Contract-Theoretic Demand Response Management in Smart Grid Systems, doi: 10.1109/ACCESS.2020.3030195, in order practically to coordinate the power consumption and needs in a power system.

 

In Section 3, the authors need to include an additional subsection discussing the computational complexity and the implementation cost of the proposed approach.

 

In Section 3, the authors need to include some additional numerical results quantifying the computational complexity of the proposed approach.

 

Finally, the overall manuscript needs to be checked for typos, syntax, and grammar errors in order to improve the quality of its presentation.

 

 

Author Response

Dear Reviewer

Regarding the review I inform you that the article has been corrected.

Author Response File: Author Response.pdf

Reviewer 4 Report

In "Forecasting of reactive power consumption with the use of artificial neural networks", the authors utilised a non-linear automated neural network to predict the course of unbalanced reactive power.  The reactive power course was registered in an industrial building and utilised as a study case for the model. The topic and approach are indeed interesting however some points should be better clarified to understand the potentialities of the work fully.

1) First of all, it is not clear the length of the dataset. In figure 9a, the measurement time is up to 5 minutes, whereas in Figure 3, the time is around 4 minutes. Is the training dataset shorter than predicted? Please specify the length of measured data and the portions selected to train and test the model. The results show that a higher prediction error is associated with the samples not included in the training.

2) The prediction error is often high than 100%. ( for time values in the [4, 5] minutes. In this case, the error with compensation is higher than the one without. Comment about this point would be appreciated.

3) It is difficult to understand the potentialities of the algorithm without having all the information about the length and composition of the training and test dataset. However, the algorithm seems to work finely only in the presence of periodic reactive power courses. Usually, this behaviour is associated with an over-fitted model. Probably it should be preferable to split data ( if the size allows it) into two distinct training and test dataset.

4) the reviewer does not entirely understand the four criteria and data reported. For example:

1. criterion I consists in determining the maximum number of consecutive samples (the so-called prediction horizon) for which the PE does not exceed 5%
2. criterion II, like criterion I, consists in determining the maximum prediction window for PE not exceeding 2%, 

Suppose the reviewer correctly understood the above statements. In that case, the second criterion seems to be more restrictive, and thus it should suppose that a lower number of samples will fulfil this criterion. Figures 8 a) and b) seem to show the opposite ( the scale of 2% is larger than the 5%). 

Also, criterion II should be less severe than criterion IV, whereas results showed the opposite.

If the reviewer correctly understands, Figure 8 c says that only two samples do not exceed 5% error in prediction. It seems not in accordance with Figure 9b).

5) if the reviewer correctly interpreted data, the results in Figures 8a), b), c) and d) are not in accordance with the results reported in Figure 8. E.g. criterion II value reported in the table is two which seems to be in accordance with panel c) rather than with panel b) as suggested in the caption.  Please check it.

The reviewer strongly suggests checking all data, tables and Figures reported. Furtermore it has to be fully explained which set of data was utilized for training and test.  

Author Response

Dear Reviewer

In attachment I'm sending answers to the questions, and at the same time I inform you, that the article has been corrected.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The topic of the paper is forecasting of 3-phase unbalanced current consumption, which is very interesting topic.  Good news is that authors use a real dataset. In this case, the data set is related to an industrial building connected to a low-voltage network, located in Gliwice city in Poland. Authors corrected the paper according the comments from the review. The paper can be published.

Reviewer 3 Report

The authors have addressed the reviewers comments.

Reviewer 4 Report

Authors successfully amended the manuscript that in the reviewer opinion is worth to be published in the current form.