A Stacking Learning Model Based on Multiple Similar Days for Short-Term Load Forecasting

Round 1

Reviewer 1 Report

The paper presents an intelligent load forecasting algorithm based on base and meta layers associated with significant data pre processing. While the topic is extremely relevant in the current energy scenario and is quite interesting for the readers there are a few questions that need to be answered. 

1. Generally from my experience in forecasting there is a sgnificant issue concerning weekdays and weekends. The load demand varies considerably during there different times. Did the authors take steps to deal with this issue ? If so, how ? 

2. The authors should refrain comparing sliding window with decomposition techniques. Decomposition techniques involve moving from the time to frequency domain or at least splitting the signal into different components in the time domain. Sliding window is a simple re-arranging of data in a way that intelligent algorithms can understand. Please correct this. 

3. Why particularly choose RBF, BPNN and RVFL ? is there a reason behind it ? 

4. Please provide the percentage improvement in the accuracy of your proposed algorithm in comparison with the best forecasting model from the comparisons. Also, please include this information in the abstract since a sentence such as "our model achieves better accuracy" is vague. 

5. Please increase the font sizes in Fig. 8 and Fig. 9, it is difficult to see the axes values 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The authors propose a hybrid predictive model, which includes a sliding window algorithm, a stacking ensemble neural network, and a similar-days predictive method.

Questions:

1. Using hybrid methods can actually improve load prediction, but the method needs to be further explained. The authors used RBF, BPNN and RVFL as base-learners, but how were the three methods used together? Were they used in cascade or in parallel? The authors used DBN and BLS-BP as meta-learners, how were these two methods used together?

2. The authors use many methods and so many variables are presented and little discussed to the point that the reader does not know their meaning. In Figure 1, for example, what is matrix H? How is it reconstructed? It would be interesting for the authors to review the article and explain the method variables.

3. In my opinion, the algorithm structure in 2.3.1 is not very clear when looking at Figures 4 and 5. Figure 5 especially is not very clear.

4. In Table 1, the authors present the parameters of the models used. Did these models have any changes for the authors' application? If these models are already present in an article, provide the reference so that the reader is able to reproduce the method proposed by the authors.

5. The comparative results are interesting, but further discussion of the results needs to be done. The results achieved by the proposed method were better than other existing methods in the literature, but even so, the errors are large. Why are the errors still large?

6. At the conclusion of the article, it would be interesting to point out the limitations of the proposed method that justify future research fields. Are there load forecasting scenarios that the proposed method does not consider?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

This paper proposes a novel ensemble learning framework for short-term load forecasting. The proposed forecasting framework employs the sliding window technique to deal with the time-series electric load data. This paper is well organized and organized. Comments to the authors:

1) What is the importance of sliding window algorithm in implementing the STLF problem?

2) What are the advantages of stacking NNs over the BPNN?

3) Similar day process requires the data to support the simulation results?

4) Explain the meta-learners in detail.

5) Provide the suitable reference for the data sets in the results section.

6) What are values of RMSE and MAPE obtained in this work?

7) What is the importance of CV folding number?

8) Provide the computations times required for each case in the results section.

9) What are the limitations of the proposed work.

10) Mention the future scope for this work?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors have answered my questions appropriately

Reviewer 2 Report

The authors propose a hybrid predictive model, which includes a sliding window algorithm, a stacking ensemble neural network, and a similar-days predictive method.

The article has been improved, the contribution is good and all questions have been effectively answered.