Wind Power Forecasting with Deep Learning Networks: Time-Series Forecasting^†

Round 1

Reviewer 1 Report

The manuscript entitled “Wind Power Forecasting with Deep Learning Networks: Time-Series Forecasting” deals with a very important topic in wind energy literature, which is 24h-72h wind power forecast.

 

The employed methods are very promising and the results look very valuable because there is also a comparison against other benchmark deep learning models.

 

In my opinion, the main flaw of this manuscript is clarity. There is a detailed description of each method and algorithm step but sincerely I have not understood the general picture, which in my opinion is essential. From what input variables does the method start and what is the output (single wind turbine power? Entire wind farm power?)? Does the method employ only SCADA data and predicts the power 24 to 72 hours later? Or is a Numerical Weather Prediction (NWP) employed? Also the SCADA data set is poorly described and in my opinion this is a flaw because the goodness of a forecast depends on the goodness of the method but also on the use of appropriate information.

 

The literature review in my opinion can be improved. The literature about 24-72 hours wind power forecast is very vast. I suggest here on some references, but the authors are encouraged to include as many as they consider appropriate.

 

• Donadio, L., Fang, J., & Porté-Agel, F. (2021). Numerical weather prediction and artificial neural network coupling for wind energy forecast. Energies, 14(2), 338.
• Hong, Y. Y., & Rioflorido, C. L. P. P. (2019). A hybrid deep learning-based neural network for 24-h ahead wind power forecasting. Applied Energy, 250, 530-539.
• Mana, M., Astolfi, D., Castellani, F., & Meißner, C. (2020). Day-ahead wind power forecast through high-resolution mesoscale model: Local computational fluid dynamics versus artificial neural network downscaling. Journal of Solar Energy Engineering, 142(3), 034502.

Author Response

The manuscript entitled “Wind Power Forecasting with Deep Learning Networks: Time-Series Forecasting” deals with a very important topic in wind energy literature, which is 24h-72h wind power forecast. The employed methods are very promising and the results look very valuable because there is also a comparison against other benchmark deep learning models.

Thanks for valuable comments. (All revised statements in manuscript were represented in brown color for easy review)

1. In my opinion, the main flaw of this manuscript is clarity. There is a detailed description of each method and algorithm step but sincerely I have not understood the general picture, which in my opinion is essential.

Response 1:

To highlight the contributions of this study for assist readers to get an overall picture, primary contributions were depicted as follows. (Sec.1)

In summary, the primary contributions of this study are as follows:

• In the presented study, the optimal parameters of model were investigated using the Evolutionary algorithms (EAs) to minimize the convergence loss of the model in the learning process.
• Four crucial architecture parameters for developed wind power prediction model are analysed incorporating with differential evolution (DE) algorithm [16–18] in the learning process of the TCN model, namely the i) number of filters, ii) activation function, iii) optimizer, and iv) dilatation coefficient to decide the initial model architecture for model training according to the natural feature of TCN.
• In our experiment, the prediction error of the TCN model for wind power prediction decreased most steadily among the four models, followed by LSTM GRU and then RNN.
• With the increasing amount of historical data, the prediction error (MAPE) of the TCN-based model decreased significantly, and the 72-h forecast error of the 1-week, 1-month, and 1-year training datasets was 66.43%, 10.93%, and 5.13%, respectively.
• Compared with LSTM GRU and then RNN models, TCN model exhibited a lower forecast error to predict 24-, 48-, and 72-h ahead of wind power generation, which is more suitable for sequence modeling based on sequence-to-sequence applications that require long effective memory, such as long wind power forecasting

2. From what input variables does the method start and what is the output (single wind turbine power? Entire wind farm power?

Response 2: Model input: real weather observations (not NWP) from Scada wind power plant in Turkey.

           Model output: multiple wind turbine power outputs in a wind farm.

 

3. Does the method employ only SCADA data and predicts the power 24 to 72 hours later? Or is a Numerical Weather Prediction (NWP) employed?

Response 3: In the training, the TCN model used only SCADA historical data and predicts the power 24 to 72 hours later. More detailed, our training dataset comprised samples from January 1, 2018, to December 26, 2018, and the test dataset used samples from 3 days, namely December 27 to 29, 2018. In real applications of project, predicting 24 to 72 hours ahead for wind power generation needs to get accurate predicted weather information from local NWP administration unit.

 

4. Also the SCADA data set is poorly described and in my opinion this is a flaw because the goodness of a forecast depends on the goodness of the method but also on the use of appropriate information.

Response 4: Supply the detailed information for Scada dataset in Sec 4.0

Scada dataset [21]

Scada Systems measure and save data's including wind speed, wind direction, generated power etc. This file was taken from a wind turbine's scada system that is working and generating power in Turkey.

The data's in the file are listed as follows:

• Date/Time: 10 minutes intervals
• LV ActivePower (kW): The power generated by the turbine for that moment.
• Wind Speed (m/s): The wind speed at the hub height of the turbine.
• Theoretical Power Curve (KWh): The theoretical power values that the turbine generates with that wind speed which is given by the turbine manufacturer.
• Wind Direction (°): The wind direction at the hub height of the turbine (wind turbines turn to this direction automatically)

5. The literature review in my opinion can be improved. The literature about 24-72 hours wind power forecast is very vast. I suggest here on some references, but the authors are encouraged to include as many as they consider appropriate.

• Donadio, L., Fang, J., & Porté-Agel, F. (2021). Numerical weather prediction and artificial neural network coupling for wind energy forecast. Energies, 14(2), 338.
• Hong, Y. Y., & Rioflorido, C. L. P. P. (2019). A hybrid deep learning-based neural network for 24-h ahead wind power forecasting. Applied Energy, 250, 530-539.
• Mana, M., Astolfi, D., Castellani, F., & Meißner, C. (2020). Day-ahead wind power forecast through high-resolution mesoscale model: Local computational fluid dynamics versus artificial neural network downscaling. Journal of Solar Energy Engineering, 142(3), 034502.

Response 5:

The technique reviews for wind power forecasting has updated with relevant and recent papers focused on the fields in the Table 1. Also, added mentioned papers as important references to References.

Reviewer 2 Report

The paper proposed a hybrid forecasting wind speed model on the convolution architecture with residual connections to learn correlations between meteorological features and wind power generation. However, the paper could not provide enough novelty and convincible results to show its superiority compared with the existing deep learning frameworks mainly due to as follows:

1. The quality of English writing is low and the paper should be revised by a native.
2. In the introduction section, the existing research gaps did not properly discuss and listed.
3. We know well "statistical and data-driven methods are more objective than the heuristic method but they need the collection of large amounts of data to produce reliable results", what is your strategy to deal with this point?
4. In order to increase the number of CNN units, the FSS formula is proposed. However, relying on just the FSS cannot guarantee to expand the CNN number with improving the quality of the classification. 
5. In the Experimental setup, some specific range of values for the hyper-parameters of CNN models are listed. However, achieving these details do not describe in the three steps of the model. These values play an important role to improve the performance of the whole CNN model. 
6. The proposed method should be compared with the state-of-the-art wind speed forecasting frameworks.
7. Developing the literature review section using some relevant methods in order to handle the hyper-parameters tuning of the deep learning model for predicting the wind speed and power can be useful such as

a) A deep learning-based evolutionary model for short-term wind speed forecasting: A case study of the Lillgrund offshore wind farm. Energy Conversion and Management, 236, 114002.

b) Multi-step-ahead wind speed forecasting based on a hybrid decomposition method and temporal convolutional networks,

Energy, 2021,121981.

c) Multi-step wind speed forecasting and Hurst analysis using novel hybrid secondary decomposition approach, Energy, Volume 238, Part A,

2022, 121764.

d) A combined forecasting system based on statistical method, artificial neural networks, and deep learning methods for short-term wind speed forecasting, Energy, Volume 217, 2021, 119361.

 

Author Response

Journal: Applied Sciences (ISSN 2076-3417)                   Ref:  applsci-1449562

Response to Reviewer 2 Comments

The paper proposed a hybrid forecasting wind speed model on the convolution architecture with residual connections to learn correlations between meteorological features and wind power generation.

Response to reviewer’s comments

Thanks for valuable comments. (All revised statements in manuscript were represented in brown color for easy review)

However, the paper could not provide enough novelty and convincible results to show its superiority compared with the existing deep learning frameworks mainly due to as follows:

1. The quality of English writing is low and the paper should be revised by a native.

 Response 1: The English writing of manuscript revision has been sent it to a native English speaker for editing service and grammar modification in my work. Minor typing mistakes and grammatical errors have corrected in revised article.

 

2. In the introduction section, the existing research gaps did not properly discuss and listed.

Response 2: Add the related discussions on recent studies for wind power predicting and highlight the gap of existing researches in Sec 1. (in brown color). These revision words are shown as follows.

 Thus, the development of an accurate and robust approach for wind power forecasting under varying climate conditions is still a challenge. Considering the increasing role of wind power in the renewable energy system, the research gaps and opportunities for wind power predicting are summarised as:

• Practically, most existing approaches to forecasting do not model uncertainty of wind well. Thus, high accuracy of wind power model needs high resolution weather data inputs generated by NWP model, which is not a trivial task.
• Typically, deep learning-based neural network for day ahead wind power forecasting outperform the traditional neural network such as ANN in the renewable power forecasting problems, due to these deep learning networks (DLNs) do not need extra data pre-processing, i.e., decomposition to retrieve the features from datasets.

 

3. We know well "statistical and data-driven methods are more objective than the heuristic method but they need the collection of large amounts of data to produce reliable results", what is your strategy to deal with this point?

Response 3: In proposed method for wind power predicting, it is a supervised machine learning model. Intrusively, the prediction performance of supervised ML techniques such as decision tree, NaiveBayes, logit regression, SVM and deep learning networks were affected by the quality of datasets. In my opinion, good quality of data can constantly assist data engineers to train the model better. In real cases, collected data were not perfect due to environment constraints including sensor aging, human operation errors,..,etc.

 For example, in experiment data of SCADA data set, some null fields contain linear proportions of the neighbouring observation data are preceded in advance. Under such constraints, some heuristic methods were developed to solve a specific problem, but these heuristic methods typically cannot produce reliable results for all cases.

 

4. In order to increase the number of CNN units, the FSS formula is proposed. However, relying on just the FSS cannot guarantee to expand the CNN number with improving the quality of the classification. 

Response 4:

In the design of TCNs for processing time-series data, the optimal parameters of the developed model shall be identified from training data to achieve high prediction precision in the model output.

In the study, the prediction error (MAPE) of the TCN-based model is affected by architecture design parameters of developed model. Thus, differential evolution (DE) algorithm is incorporated to search the optimal design parameters of developed model. Typically, DE have provided good search results to minimize the prediction error. In the experiment, TCN model exhibits the consistent, stable and good prediction results by using selected parameters from DE algorithm. 

 

5. In the Experimental setup, some specific range of values for the hyper-parameters of CNN models are listed. However, achieving these details do not describe in the three steps of the model. These values play an important role to improve the performance of the whole CNN model. 

Response 5:

The original parameters of developed TCN models are derived from P. Rémy at GitHub which was revised as follows.

In this step, four crucial architecture parameters were selected from transferring learning cases in the TCN predictor [20-23] and the original parameters of developed TCN models are obtained from P. Rémy at GitHub [29] for developed wind power prediction model

 

6. The proposed method should be compared with the state-of-the-art wind speed forecasting frameworks.

Response 6: The proposed method is compared with accuracy of three real projects for wind speed forecasting in the European from 2011~2019 and described as follows. (Sec.4)

In summary, the wind power forecast error (MRE) of proposed TCN-based model is near 5.13% based on 1-year historical data in different climatic scenarios; Compared to accuracy other projects in wind power forecasting, the European team’s SafeWind project in 2011 achieved the forecasting error of 17%. In 2017, the predicted error improved to 11%, and then the project of BSI electric power company reached within 10% in 2019; As shown in Table 11, the proposed TCN-based approach provides a lower predicted error with higher prediction accuracy than those of real projects in studies of wind power forecasting [31–33].

 

7. Developing the literature review section using some relevant methods in order to handle the hyper-parameters tuning of the deep learning model for predicting the wind speed and power can be useful such as

a) A deep learning-based evolutionary model for short-term wind speed forecasting: A case study of the Lillgrund offshore wind farm. Energy Conversion and Management, 236, 114002.

b) Multi-step-ahead wind speed forecasting based on a hybrid decomposition method and temporal convolutional networks, Energy, 2021,121981.

c) Multi-step wind speed forecasting and Hurst analysis using novel hybrid secondary decomposition approach, Energy, Volume 238, Part A, 2022, 121764.

d) A combined forecasting system based on statistical method, artificial neural networks, and deep learning methods for short-term wind speed forecasting, Energy, Volume 217, 2021, 119361.

Response 7:

The technique reviews for wind power forecasting has updated with relevant and recent papers focused on the fields in the Table 1. Also, added mentioned papers for hyper- parameters tuning of the deep learning model in our manuscript as important references to References.

Reviewer 3 Report

The work titled Wind Power Forecasting with Deep Learning Networks: Time-Series Forecasting, discusses a key topic of interest in the wind industry. The authors propose their methodology and provide evidence of its better than other methods. Some observations:

• The methodology used is ok however one main concern is the complexity of the functions and implementation of the analysis. Guess it would need some further refinement to make it to the operators.
• It may be better if the literature review is confined to one section and is made distinct from the other section.    

Overall the work is decent and flows well with a good layout of scientific material and analysis.  

Author Response

Journal: Applied Sciences (ISSN 2076-3417)                   Ref:  applsci-1449562

Response to Reviewer 3 Comments

The work titled Wind Power Forecasting with Deep Learning Networks: Time-Series Forecasting, discusses a key topic of interest in the wind industry.

Thanks for valuable comments. (All revised statements in manuscript were represented in brown color for easy review)

The authors propose their methodology and provide evidence of its better than other methods.

Response 1: The proposed method is compared with accuracy of three real projects for wind speed forecasting in the European from 2011~2019 and described as follows. (Sec.4)

In summary, the wind power forecast error (MRE) of proposed TCN-based model is near 5.13% based on 1-year historical data in different climatic scenarios; Compared to accuracy other projects in wind power forecasting, the European team’s SafeWind project in 2011 achieved the forecasting error of 17%. In 2017, the predicted error improved to 11%, and then the project of BSI electric power company reached within 10% in 2019; As shown in Table 11, the proposed TCN-based approach provides a lower predicted error with higher prediction accuracy than those of real projects in studies of wind power forecasting [31–33].

2.The methodology used is ok however one main concern is the complexity of the functions and implementation of the analysis. Guess it would need some further refinement to make it to the operators.

Response 2: Thanks for valuable comments. To improve readability of proposed model, we revised the context of the implementation and the analysis step by step according to Figure 10 Experiment execution process. In other words, steps of implementation and analysis in Sec 3.3 and Sec 4.1 are followed the steps presented in Figure 10. Also descriptions of regarding differential evolution in Sec 3.2.2. Analysis of architecture design were merged to Sec 2.2.

 

• It may be better if the literature review is confined to one section and is made distinct from the other section.  Overall the work is decent and flows well with a good layout of scientific material and analysis.

Response 3:Thanks for valuable comments. We move parts of content (including Table 1) to Section 2 and rename Section two as 2.literature review for clarity.

Round 2

Reviewer 1 Report

The authors have addressed my comments. The paper in my opinion is adequate for publication.

Reviewer 2 Report

The authors have sufficiently addressed the reviewed issues in the manuscript.