the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 13 Jun 2023
Forecasting of wind power by using a hybrid machine learning method for the Nord-Pool intraday electricity market
Abstract. The interest in trading intraday markets has been increasing due to the growth of renewable intermittent energy production. With the growing renewable energy capacity, which mostly comes from wind energy, the intraday market volume has been continuously increasing every year. In Europe, countries work with different lead times ranging from 5 to 90 minutes and trading blocks of 15 minutes. Several countries, including Sweden, use 15-minute trading blocks with 60 minutes lead time. Market participants use the intraday market to optimize their position after the day-ahead market closes. Since new methods become available, such as better forecasts on short-term renewable energy power output and demand, the intraday market has become more important for energy traders in order to maximize their profit. The primary objective of this study is to enhance the intraday forecasting of wind power by improving the forecasting methods using machine learning. A hybrid approach that combines a mode decomposition method, Empirical Mode Decomposition (EMD), with Support Vector Regression (SVR), is used. In addition, the forecasting with the SVR method is improved by applying a cross-validation method that tunes the parameters used. The study utilized three months (92 days) of wind turbine power data from 21 June 2017 to 20 September 2017. 80 % of the data was used for training, and the remaining data were used for predictions. The results showed that combining SVR with a hybrid method that incorporates EMD predictions can lead to higher prediction accuracy. Furthermore, our results stress that parameter-tuning algorithms can improve machine-learning methods. We believe that the methods proposed in this study will be beneficial for the planning of dispatchable energy generation and pricing for the intraday electricity market.
This preprint has been withdrawn.
-
Withdrawal notice
This preprint has been withdrawn.
-
Preprint
(1054 KB)
Interactive discussion
Status: closed
-
RC1: 'Comment on wes-2023-48', Anonymous Referee #1, 21 Jun 2023
The paper concentrates on a timely and relevant problem, i.e., forecasting for the intra-day market participation of wind power generation assets. The literature on wind power forecasting is vast already (especially with market participation applications in mind), with a wealth of models, methods, case studies, etc. Hence, new papers on the topic need to bring significant novelty to describe new approaches, new challenges and solutions, etc. In the present case, it is not clear what the novelty is since
1. using decomposition and then some ML machinery (including SVR) is definitely not new. It has been done numerous times for renewable energy forecasting, in many different configurations (e.g., different approaches to decomposition, different ML techniques, etc.)
2. using cross validation, and concluding that parameter tuning techniques are beneficial, is an obvious results. This is what is taught in "Intro to ML" courses at various universities, and is part of any best practice for application of ML approaches to renewable energy forecasting problem.
In parallel, one can identify potentially key issues that would need to be clarified before the paper could be considered for publication. These relate to potential data leakage. In the way the work is presented, I suspect that the data decomposition was done for all data at once (so, for train, CV, test data), which may introduce data leakage. Either the pipeline should be adapted, or the authors should demonstrate there is no data leakage.
Finally, the approach to forecast verification is very light and not necessarily rigorous. Only 3 months of data is used here overall, which is extremely short. The authors eventually end up with 16 days for genuine forecast verification. This makes it difficult to yield conclusions and insights that would be of general value in terms of forecasting. There is also no real benchmark that would allow to place the performance of the approach in perspective with the exising. Hence, the aim is only to show that decomposition and cross-validation may be beneficial if using SVR. Back to the lack of novelty issue, in all cases, these results are already available in the literature.
Citation: https://doi.org/10.5194/wes-2023-48-RC1 -
RC2: 'Comment on wes-2023-48', Anonymous Referee #2, 26 Jul 2023
In the paper “Forecasting of wind power by using a hybrid machine learning method for the Nord-Pool intraday electricity market”, the authors use Support Vector Regression (with and without decomposition) in forecasting wind generation. The use of sub-hourly resolution and its relation to intra-day markets is interesting, but the authors do not sufficiently justify the model selection and the model validation is insufficient (only 1 location and only a few months of data). I find the literature review insufficient, and the omission of using numerical weather prediction data is not justified. The novelty of the paper is not clear. The proposed method should be validated against state-of-the-art forecasting techniques (other than SVR).
1) The literature review is insufficient. In addition to listing some forecasting techniques, the authors should describe which methods have been found to be the best in the literature for wind forecasting. Especially, has the chosen method been shown to outperform other methods in wind generation forecasting? It seems to me that the authors do not discuss all relevant forecasting techniques: e.g., deep learning or random forest type (decision trees/ensemble learning) methods are not really covered. Why the use of numerical weather prediction data as part of the forecasting process (as is often done) is not discussed?
2) The authors say: “However, few researchers have paid attention to the forecasting problem about the wind power capacity, which plays an important role in wind power construction plans, investment, operation, and as well as energy trading plan”. What is the forecasting problem about the wind power capacity and has it been solved in this paper?
3) The line “The forecasting methods include autoregressive models, ARIMA models, artificial neural network models, and support vector regression models” is repeated almost the same around row 35 and row 40. Why is it mentioned twice (I don’t understand the difference of things discussed around line 35 and around line 40)?
4) The authors write: “Recently, hybrid methods have promised improvements in short-term wind power forecasting and uncertainty analysis”. What are the hybrid methods, and why they shown improvement (and improvement compared to what models)?
5) The authors write: “The results imply that combining SVR with a hybrid method that incorporates EMD predictions can lead to higher prediction accuracy”. Higher compared to what?
6) For what reason is only 11.00 to 17.00 data used? Why not use all hours of the day?
7) The authors write: “There are missing records, meaning that for some seconds the turbine has generated no power, which are excluded.” Why were time steps where the turbine did not generate excluded? Is it not important information (if the generation was zero for some time step)?
8) Why data from only 1 wind turbine is used? It would be much more relevant to forecast the generation of the entire power plant.
9) Why only 3 months of data are used? This is in my view much too few data points to draw any significant conclusions (at least a few years of data would be preferable). Also, the use of only 1 location is very limited in my view, as one would like to compare at least a few locations to see if the results are robust.
10) Have the authors considered using numerical weather prediction data as a part of their model? When forecasting more than 1 hour ahead (as is the case in this paper), the combination of weather forecasts and forecasts based on previous measurements can perform better than using only the previous measurements. The authors should test and compare using numerical weather prediction as part of their model to the current approach.
11) The authors list many forecast techniques in the literature review (e.g., ARIMA and neural networks), but the chosen approach is not compared to other possible approaches. Why did the authors choose to use support vector machine (SVM) regression and not, e.g., neural networks? I would also expect to authors to compare SVM to at least 1 other often used forecast technique (e.g., random forest) to put the proposed methodology in perspective (comparison of the decomposition technique is not sufficient in my view, as the method is still fundamentally SVM).
12) Why the specific decomposition technique was used? Are there other possible decomposition techniques and if yes, why the authors consider the selected method to be the most appropriate?
13) Why RMSE was used as the metric to compare to forecast results? Are there other possible metrics, and if yes, why use specifically RMSE?
14) The authors say that: “It should be noted that all parameters for both SVR and EMD are kept the same for all predictions”. Is this necessarily the most appropriate way? Could there be any benefit in dynamically updating/recalibrating the parameters?
Citation: https://doi.org/10.5194/wes-2023-48-RC2 -
EC1: 'Comment on wes-2023-48', Nikolay Dimitrov, 08 Sep 2023
Dear Authors,
Thank you for choosing the Wind Energy Science journal for submitting your manuscript. The paper has now been reviewed by two experts in the field. The reviewers are unfortunately raising some significant issues with the paper. As we need to abide by the recommendations of the experts, I have no other choice but to recommend rejection of the current version of the manuscript. You are more than welcome to consider a new submission at a later stage when you have addressed the current issues. We will be happy to consider it again for publication in WES.
Best regards,
Nikolay Dimitrov, Associate Editor
Citation: https://doi.org/10.5194/wes-2023-48-EC1
Interactive discussion
Status: closed
-
RC1: 'Comment on wes-2023-48', Anonymous Referee #1, 21 Jun 2023
The paper concentrates on a timely and relevant problem, i.e., forecasting for the intra-day market participation of wind power generation assets. The literature on wind power forecasting is vast already (especially with market participation applications in mind), with a wealth of models, methods, case studies, etc. Hence, new papers on the topic need to bring significant novelty to describe new approaches, new challenges and solutions, etc. In the present case, it is not clear what the novelty is since
1. using decomposition and then some ML machinery (including SVR) is definitely not new. It has been done numerous times for renewable energy forecasting, in many different configurations (e.g., different approaches to decomposition, different ML techniques, etc.)
2. using cross validation, and concluding that parameter tuning techniques are beneficial, is an obvious results. This is what is taught in "Intro to ML" courses at various universities, and is part of any best practice for application of ML approaches to renewable energy forecasting problem.
In parallel, one can identify potentially key issues that would need to be clarified before the paper could be considered for publication. These relate to potential data leakage. In the way the work is presented, I suspect that the data decomposition was done for all data at once (so, for train, CV, test data), which may introduce data leakage. Either the pipeline should be adapted, or the authors should demonstrate there is no data leakage.
Finally, the approach to forecast verification is very light and not necessarily rigorous. Only 3 months of data is used here overall, which is extremely short. The authors eventually end up with 16 days for genuine forecast verification. This makes it difficult to yield conclusions and insights that would be of general value in terms of forecasting. There is also no real benchmark that would allow to place the performance of the approach in perspective with the exising. Hence, the aim is only to show that decomposition and cross-validation may be beneficial if using SVR. Back to the lack of novelty issue, in all cases, these results are already available in the literature.
Citation: https://doi.org/10.5194/wes-2023-48-RC1 -
RC2: 'Comment on wes-2023-48', Anonymous Referee #2, 26 Jul 2023
In the paper “Forecasting of wind power by using a hybrid machine learning method for the Nord-Pool intraday electricity market”, the authors use Support Vector Regression (with and without decomposition) in forecasting wind generation. The use of sub-hourly resolution and its relation to intra-day markets is interesting, but the authors do not sufficiently justify the model selection and the model validation is insufficient (only 1 location and only a few months of data). I find the literature review insufficient, and the omission of using numerical weather prediction data is not justified. The novelty of the paper is not clear. The proposed method should be validated against state-of-the-art forecasting techniques (other than SVR).
1) The literature review is insufficient. In addition to listing some forecasting techniques, the authors should describe which methods have been found to be the best in the literature for wind forecasting. Especially, has the chosen method been shown to outperform other methods in wind generation forecasting? It seems to me that the authors do not discuss all relevant forecasting techniques: e.g., deep learning or random forest type (decision trees/ensemble learning) methods are not really covered. Why the use of numerical weather prediction data as part of the forecasting process (as is often done) is not discussed?
2) The authors say: “However, few researchers have paid attention to the forecasting problem about the wind power capacity, which plays an important role in wind power construction plans, investment, operation, and as well as energy trading plan”. What is the forecasting problem about the wind power capacity and has it been solved in this paper?
3) The line “The forecasting methods include autoregressive models, ARIMA models, artificial neural network models, and support vector regression models” is repeated almost the same around row 35 and row 40. Why is it mentioned twice (I don’t understand the difference of things discussed around line 35 and around line 40)?
4) The authors write: “Recently, hybrid methods have promised improvements in short-term wind power forecasting and uncertainty analysis”. What are the hybrid methods, and why they shown improvement (and improvement compared to what models)?
5) The authors write: “The results imply that combining SVR with a hybrid method that incorporates EMD predictions can lead to higher prediction accuracy”. Higher compared to what?
6) For what reason is only 11.00 to 17.00 data used? Why not use all hours of the day?
7) The authors write: “There are missing records, meaning that for some seconds the turbine has generated no power, which are excluded.” Why were time steps where the turbine did not generate excluded? Is it not important information (if the generation was zero for some time step)?
8) Why data from only 1 wind turbine is used? It would be much more relevant to forecast the generation of the entire power plant.
9) Why only 3 months of data are used? This is in my view much too few data points to draw any significant conclusions (at least a few years of data would be preferable). Also, the use of only 1 location is very limited in my view, as one would like to compare at least a few locations to see if the results are robust.
10) Have the authors considered using numerical weather prediction data as a part of their model? When forecasting more than 1 hour ahead (as is the case in this paper), the combination of weather forecasts and forecasts based on previous measurements can perform better than using only the previous measurements. The authors should test and compare using numerical weather prediction as part of their model to the current approach.
11) The authors list many forecast techniques in the literature review (e.g., ARIMA and neural networks), but the chosen approach is not compared to other possible approaches. Why did the authors choose to use support vector machine (SVM) regression and not, e.g., neural networks? I would also expect to authors to compare SVM to at least 1 other often used forecast technique (e.g., random forest) to put the proposed methodology in perspective (comparison of the decomposition technique is not sufficient in my view, as the method is still fundamentally SVM).
12) Why the specific decomposition technique was used? Are there other possible decomposition techniques and if yes, why the authors consider the selected method to be the most appropriate?
13) Why RMSE was used as the metric to compare to forecast results? Are there other possible metrics, and if yes, why use specifically RMSE?
14) The authors say that: “It should be noted that all parameters for both SVR and EMD are kept the same for all predictions”. Is this necessarily the most appropriate way? Could there be any benefit in dynamically updating/recalibrating the parameters?
Citation: https://doi.org/10.5194/wes-2023-48-RC2 -
EC1: 'Comment on wes-2023-48', Nikolay Dimitrov, 08 Sep 2023
Dear Authors,
Thank you for choosing the Wind Energy Science journal for submitting your manuscript. The paper has now been reviewed by two experts in the field. The reviewers are unfortunately raising some significant issues with the paper. As we need to abide by the recommendations of the experts, I have no other choice but to recommend rejection of the current version of the manuscript. You are more than welcome to consider a new submission at a later stage when you have addressed the current issues. We will be happy to consider it again for publication in WES.
Best regards,
Nikolay Dimitrov, Associate Editor
Citation: https://doi.org/10.5194/wes-2023-48-EC1
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 561 | 318 | 33 | 912 | 22 | 22 |
- HTML: 561
- PDF: 318
- XML: 33
- Total: 912
- BibTeX: 22
- EndNote: 22
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
