An Improved Residential Electricity Load Forecasting Using a Machine-Learning-Based Feature Selection Approach and a Proposed Integration Strategy

Round 1

Reviewer 1 Report

The authors proposed a load forecasting framework with time-series autoregressive approaches, BGA-PCA feature selection, and a feed-forward neural network. The authors intended to improve the forecasting performance by selecting a subset of features and minimizing the MAPE between the original household load sequence and the forecasted sequence.

 

However, this paper can be improved in the following aspects. 

• The authors’ understanding of the concept of “artificial intelligence (AI)” demonstrated in the second paragraph of Section 2 “Related Work” might not be comprehensive enough. Artificial intelligence (AI) is a very broad subject and is more than just neural networks (NN) and support vector machines (SVM).
• The authors reviewed a few previous studies on machine learning-based load forecasting in the first paragraph of Section 3.2, yet did not cite any corresponding references. Here, it might be better to include citations when reviewing existing works.
• When introducing the architectures of machine learning models, the definition of loss function and training schemes are generally the two key pieces of information to provide. However, these details were not presented in this manuscript. Also, how exactly the time series need to be preprocessed in order to fit in for the machine learning model, and how exactly this machine learning model has been used to “validate” the proposed autoregressive model are not quite clear.
• The authors could also consider using figures to illustrate the model structure (in this case, the ANFIS model) instead of using only written descriptions, which might provide the readers with a clearer or more intuitive grasp of the subject.
• Grammatical mistakes or confounding sentences detected:
  • At lines 58-59: “The aggregator depends upon the enabling technologies such that controllable loads, micro-generators, and distributed energy resources” – “as”
  • At lines 420-422: “Here is presented a new time series and autoregression algorithm for analyzing the more complex and massive historical data” – “we”
  • At lines 422-423: “The model is further validated by ML-based feedback ANN model of five stages is computed for the next step of feature selection” – Not grammatically correct
  • At lines 423-425: “Our model of best feature selection BGA-PCA, given a best feature subset than the actual data with initial features” – Confounding

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

The article investigates the demand-side load forecasting problem by introducing a workflow with machine learning methods. A case study is provided using actual power consumption data. The research idea is original, innovate and interesting, but the presentation quality must be improved.

The authors should explain why the reviewed literatures are related (differences and relations to this paper), rather than just present what these works do in Section 2. In Table 1, the attribute "Machine Learning/Deep learning" is confusing and not necessary in my opinion.

The proposed framework is not clearly presented. The authors mentioned Phase 2 but there is no Phase 1, and no description of the contained tasks in these phases. The presentations of the aggregation method, the feedforward ANN, and the feature selection method are not well organized. The authors should at least provide the input/output of these methods, and how these methods work together.

Figure 6 is redundant. The authors already provide some metrics to evaluate their methods and show the plots of four seasons.

On Line 186, there should be a citation.

 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

General comment

This paper presents a load forecasting model based on Machine Learning approach. The topic presented is of interest to the readership of this journal. Introduction and Related work sections accurately report the objectives of the paper and the literature on the topic, with in-depth information on the different approaches with which the forecasts were made. The results are also clearly described, with the figures that can be partially improved to make them more understandable. I recommend a decision of minor revisions. Below are reported some minor comments.

 

Minor comments

• Line 19: …, respectively à missing comma
• Line 19: “The algorithm of seven different autoregression models is also computed and we analyzed a Feedforward Artificial Neural Network (FANN) model for further validation based on the ML approach.” à An algorithm, which consists of seven different autoregression models, has been also developed and validated through a Feedforward Artificial Neural Network (FANN) model, based on the ML approach.
• Line 26: of MAPE, equal to70%... and spring seasons, respectively.
• Lines 199, 200, 201, 202, 213: check the symbols, in particular superscript and subscript fonts.
• Figures from 2 to 6: it would be advisable to represent axes in black and not in red.
• Figure 4 on the right: there is a typo, I think it is the Matlab selection tool.

 

 

 

 

 

 

• Figures 2 to 5: it would be better to report the error graph below the comparison graph between measured and predicted, in this way the comparison becomes easier.

As a template example:

In your case you could report in the graph above only "Actual" vs "Model" in red and blue as you have already done, and below the error bar graph in orange. The axes in black.

 

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.pdf