Transformer-Based Hybrid Forecasting Model for Multivariate Renewable Energy

Round 1

Reviewer 1 Report

This paper presents a hybrid model combining a statistical method (SARIMA) with several ML algorithms for the prediction of energy produced by renewable energy stations (solar and wind).

The authors contextualise the problem in the difficulty of calculating the energy produced by these plants and continue the line of exploration followed in the literature with the proposal of new combinations of linear and non-linear models for regression. 

The paper is very well written and organised, with a clear explanation of the proposed model and results that demonstrate an improvement in the area. My suggestion is that the paper can be published after improving several aspects that can be considered minor:

- The authors should make clearer the contribution of their work to the renewable energy industry, highlighting their contribution from the point of view of the benefits for this industry.

- Add more evaluation metrics, such as Coefficient of Variation of Root-Mean Squared Error - CV(RMSE).

- Clearly justify why the SARIMA model has been chosen for the linear module and not another model.

- Is it possible to add other exogenous variables to the input data, such as the effective production capacity of each solar panel?

- Figures 3 and 4 should show negative gain values on the X-axis.

- In the conclusions the authors should mention the limitations of their work.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In this paper, the authors proposed a transformer-based hybrid forecasting model for multivariate renewable energy. This method combines the traditional linear model with a state-of-the-art machine learning model and transformers neural networks. Although the authors have compared with the existing methods, the proposed method in this paper is not detailed enough, such as the network structure, the whole training and test processes. In addition, there is some confusion in the organization of this paper. And there are also some unreasonable expressions in this paper, such as some mistakes in Table 1.  So, this paper cannot meet the requirements for publication in this journal.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Presented approach of model reidual series estimation/forecasting is known for many years now. Authors used a "Transformer" network which is a kind of novelty in this context. 

Remarks:

1. Formula (7) MSE - missing square symbol

2. Formula (10) - there is no exact definition or determination what is "Metric_literature". What was the method or maybe methods. What were the values taken to calculate "Gain". Without this Figure 3 and 4 are hard to interpret and useless.

3. Table 3 - there is an enormous difference between "Solar" and "Wind" error metrics values. Probably, the "Solar" timeseries were not normalised. If so RMSE and MAE should be expressed in kW or MW (kWh, MWh). Comparing the quality of forecasts for "Solar" and "Wind" is very difficult. In other words, the most important table in the article is unreadable.

4. The above remark concerns figure 5 also.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

The paper proposes an interesting novel concept for the hybrid approach, which connects the SARIMA method with standard machine learning methodologies.

The introduction of the paper shows the importance of creating new forecaster models. However, the following paper is not discussed in the Introduction doi.org/10.1016/S0925-2312(01)00702-0, which also introduced an H-transformer, its different from the topic of the paper because it uses ARIMA and not SARIMA, but the advantage of this different is must be highlighted to the reader. 

From the other point of view, the topic and motivation of the paper is pretty good and interesting.  The paper shows a comparison with different ml methodologies. This comparison is interesting and the Wolpert's no-free-lunch theorem of the mathematical optimization/machine learning can be discussed here because its supports the paper.  The main strength this statement (https://doi.org/10.3390/app10196653, section 2), which states that there is not existing best method which can outperform all of the optimization solvers in all kinds of problems.

The results and discussion section should be improved because only about 15 lines long, it is very small. There are many interesting differences and details should be discussed here, moreover, the used codes and the test of the codes can be described. Is the used code is openly available in a repository?

It would be nice of the conclusion and the future works subsection is merged into one chapter, with very minimal number of citations to share a more understandable "takeway" for the reader.  I think this part can be merged and justify the novelty of the paper at the end of the introduction.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 4 Report

The authors answered all of my questions.