Novel PV Power Hybrid Prediction Model Based on FL Co-Training Method

Round 1

Reviewer 1 Report

Les authors must complete several parts:

1)     What is the data acquisition flow chart for this study?

2)     What is the architecture of the training model?

3)     How much data and what type of data have been studied and used?

4)     In which real case these data and this method were used (smart grids, microgrids …)?

5)     What is the max training epochs number?

6)     What is the number of hidden units?

7)     What is the accuracy percentage of this method compared to other methods in the literature?

8)     The authors can use the articles below to improve their paper:      

-   A Hybrid Photovoltaic Power Prediction Model Based on Multi-source Data Fusion and Deep Learning, IEEE-SCEMS 2020

-    A 3D Design of Small Hybrid Farm for Microgrids, IEEE-WAC2022

-        A hybrid machine learning method with explicit time encoding for improved Malaysian photovoltaic power prediction, ELSEVIER-Journal of Cleaner Production , 2022

Author Response

(1) What is the data acquisition flow chart for this study?

Response: As shown in Figure 5, we have added the data acquisition process of our work. First, the photovoltaic data collector (smart meter) collects the current, voltage and other effective data from each photovoltaic user. Then, the local communication HPLC （High-speed power line carrier）uploads the data collected by each smart meter to the local data acquisition terminal. Finally, the local data is transmitted to the main station through 4G network communication, that is, the electricity information collection system. The additions are marked in red.

(2) What is the architecture of the training model? 

Response: Firstly, the model training in this paper is divided into two parts: (1) local model training. (2) Cloud model parameter aggregation training. As shown in Figure 1, this paper adopts a cloud-edge architecture, which contains a cloud center and many edges. The cloud carries the FL algorithm that aggregates the model parameters of each region, and the edges are the local server of each participant, which carry the prediction algorithm. The local data is used to train the prediction model locally, but due to the limitations of local data, it is necessary to upload the local model parameters to the cloud for parameter aggregation training again. The specific model training process has been described in detail in the Overall Program section, and is also marked in red.

(3) How much data and what type of data have been studied and used?

Response: The research data set contains photovoltaic positive active energy data, active power, voltage, current and date information recorded every 15 minutes in 20 regions. Five regions were selected as independent photovoltaic power generation companies to participate in the federal study. The measured data of the five regions from June to November 2022 were selected, and recorded once every 1h, with more than 4000 data samples of each region. In order to verify the predictive ability of this method, this paper uses the data samples from June 1 to November 20 as the model training sample set, and 24 data samples from the data in November 21 as the test set. The corresponding content has been added and marked in red.

(4) In which real case these data and this method were used (smart grids, microgrids …)?

Response: As shown in Experiment Results and Discussions Section, the method proposed in this paper has been experimented in the grid system of Hebei Province, China. The data are from distributed PV customers in five regions in Hebei Province, China. The prediction model has also been tested in these five regions.

(5) What is the max training epochs number? 

Response: The maximum number of cycles for local model training set in this experiment is 6000, and the maximum number of cycles for FL model aggregation is 1000. We add these contents to the experiment part of the article and mark them in red.

(6) What is the number of hidden units?

Response: The main cycle of the data is 24 hours, so the time step of the mixed model is set to 24, and the number of cell units described in the Overview of the prediction model algorithm section is 24. As for the BP neural network (BPNN), the optimal value of hidden layer is 1, which means the number of hidden units is 24. The corresponding content has been added and marked in red.

(7) What is the accuracy percentage of this method compared to other methods in the literature?

Response: In this paper, the values of RMSE (Root Mean Squared Error) and MAPE (Mean Absolute Percentage Error) are used to represent the accuracy. The smaller the index value, the smaller the error between the predicted value and the true value, and the better the prediction effect of the model. At the same time, in order to reflect the advantages of the improved prediction model in this paper, comparison of prediction model indicators is added: compare the prediction algorithm in this paper with the traditional BPNN and LSTM, as shown in Table 3. Compared with BPNN and LSTM, the proposed hybrid prediction model LSTM-BPNN has the smallest MAPE value of 2.63%. Therefore, the hybrid LSTM-BPNN model proposed in this paper possesses more accurate prediction results than the two original models.

The added content is added in the experiment section and marked in red. 

(8) The authors can use the articles below to improve their paper:      

-   A Hybrid Photovoltaic Power Prediction Model Based on Multi-source Data Fusion and Deep Learning, IEEE-SCEMS 2020

-    A 3D Design of Small Hybrid Farm for Microgrids, IEEE-WAC2022

-        A hybrid machine learning method with explicit time encoding for improved Malaysian photovoltaic power prediction, ELSEVIER-Journal of Cleaner Production , 2022

Response: Thank you very much for your suggestions. We highly recognize the three articles you recommended and have added them to the references[13], [14], [15]. We believe that the level of our references can be improved. 

Reviewer 2 Report

The manuscript needs following revisions to become acceptable for publication:

1. It is better to quantitatively assess the model and represent the results in abstract section. 

2. There are some writing errors. It is better to recheck text and edit them. 

3. Importance and novelties of the work should be highlighted. 

4. Quality of the figures is very good and attractive. 

5. Adding nomenclature is suggested. 

6. It is better to use following references to improve literature review in term of applications of PV modules, influential factors and intelligent methods:

"Techno-Economic Analysis and Optimization of an Off-Grid Hybrid Photovoltaic–Diesel–Battery System: Effect of Solar Tracker" https://doi.org/10.3390/su14127296

"PV/Thermal as Promising Technologies in Buildings: A Comprehensive Review on Exergy Analysis" https://doi.org/10.3390/su141912298

"Thermal Management of Solar Photovoltaic Cell by Using Single Walled Carbon Nanotube (SWCNT)/Water: Numerical Simulation and Sensitivity Analysis" https://doi.org/10.3390/su141811523

"Cooling of Photovoltaic Panel Equipped with Single Circular Heat Pipe: an Experimental Study" https://doi.org/10.22044/rera.2022.11523.1097

7. Adding some relevant references from electronics journal can be useful. 

8. It is better to add some quantitative results in conclusion section. 

Author Response

(1) It is better to quantitatively assess the model and represent the results in abstract section.  

Response: Thank you for your suggestion. We add quantitative assess of the experimental results in Abstract section. The experimental results show that the minimum MAPE of the hybrid prediction model constructed in this paper can reach 1.2%, and the prediction effect is improved by 30% compared with the traditional model. Under the FL mode, the trained prediction model not only improves the prediction accuracy by more than 20%, but also has excellent generalization ability in multiple scenarios. The revision part is marked in red.

(2) There are some writing errors. It is better to recheck text and edit them. 

Response: Thank you for your correction of the article. We checked the full text again and corrected the errors.

(3) Importance and novelties of the work should be highlighted.

Response: According to your suggestion, we have strengthened the elaboration of the importance and novelties of this work. We provide a description of the importance and novelties in the introduction section, highlighting the new prediction method proposed in this paper. The specific contents are as follows, and the corresponding parts are also marked in red.

• This paper proposes a hybrid prediction model. The model consists of LSTM and BPNN. The LSTM is used to extract important features from the time-series data, and the BPNN can compensate for the shortcomings of the LSTM network's insufficient fitting ability to achieve higher accuracy energy consumption prediction.
• This paper is the first to propose a FL-LSTM-BPNN model for PV power prediction. The hybrid prediction model is trained collaboratively under FL, and the data features of each company are federated. It can both improve the model generalization ability, reduce the communication cost, and protect the data privacy.

(4) Quality of the figures is very good and attractive.

Response: Thank you very much for your recognition of our work.

(5) Adding nomenclature is suggested.

Response: Thank you very much for your suggestion, the method proposed in this paper consists of a combination of FL and LSTM-BPNN models, hence the name FL-LSTM-BPNN. and we have included this within the paper, which is marked in red.

(6) It is better to use following references to improve literature review in term of applications of PV modules, influential factors and intelligent methods:

"Techno-Economic Analysis and Optimization of an Off-Grid Hybrid Photovoltaic–Diesel–Battery System: Effect of Solar Tracker" https://doi.org/10.3390/su14127296

"PV/Thermal as Promising Technologies in Buildings: A Comprehensive Review on Exergy Analysis" https://doi.org/10.3390/su141912298

"Thermal Management of Solar Photovoltaic Cell by Using Single Walled Carbon Nanotube (SWCNT)/Water: Numerical Simulation and Sensitivity Analysis" https://doi.org/10.3390/su141811523

"Cooling of Photovoltaic Panel Equipped with Single Circular Heat Pipe: an Experimental Study" https://doi.org/10.22044/rera.2022.11523.1097.

Response: Thank you very much for your suggestions. We highly recognize the four articles you recommended and have added two of them to the references[8], [9]. We believe that the level of our references can be improved.

(7) Adding some relevant references from electronics journal can be useful.

Response: Thank you for your advice. According to your advice, some of the references in the article are from electronic journal such as [2], [4], [17], [24].

(8) It is better to add some quantitative results in conclusion section.

Response: According to your suggestion, we have edited the conclusion part again and added the experimental results to directly explain the advantages of our work. The specific changes are as follows. Finally, through quantitative experiments, the MAPE of the hybrid LSTM-BPNN prediction model proposed in this paper is 2.63%, which is a 20% improvement over the traditional LSTM model. In a large number of experiments under the collaborative model training mode of FL, the prediction performance indexes RMSE and MAPE of the model can be reduced by 60% , which verifies the improvement of the prediction accuracy of the model by the proposed method in this paper, and proves that the short-term load prediction model trained by the method in this paper has excellent generalization ability in multiple scenarios. The specific changes are marked in red in the paper.

Reviewer 3 Report

The paper proposes a method for PV power production. Although the authors claimed that this is a novel method, the results did not show that as the accuracy of the results is low as compared to similar methods including the conventional method. The utilized data is also not defined well whereas the step is no well defined. Here the authors are invited to comments on this issue and compare their work in a serious way with previous LSTM models 

Author Response

The paper proposes a method for PV power production. Although the authors claimed that this is a novel method, the results did not show that as the accuracy of the results is low as compared to similar methods including the conventional method. The utilized data is also not defined well whereas the step is no well defined. Here the authors are invited to comments on this issue and compare their work in a serious way with previous LSTM models

Response: Thank you for taking your valuable time to correct our work. We have made changes to the article based on your suggestions, and the changes are as follows.

• We have added a comparison of the predictive metrics of the method in this paper with the traditional BPNN and LSTM algorithms. The commonly used values of RMSE (root mean square error) and MAPE (mean absolute percentage error) are used in this paper to express accuracy. The results are shown in Table3, Compared with BPNN and LSTM, the experimental results show that the proposed hybrid prediction model LSTM-BPNN has the smallest MAPE value of 2.63% . Therefore, the hybrid LSTM-BPNN model proposed in this paper possesses more accurate prediction results than the two original models. The additions are shown below and have been marked in red throughout the paper. 

• We have provided better explanatory notes on the data used in this paper. The research data set contains photovoltaic positive active energy data, active power, voltage, current and date information recorded every 15 minutes in 20 regions. Five regions were selected as independent photovoltaic power generation companies to participate in the federal study. The measured data of the five regions from June to November 2022 were selected, and recorded once every 1h, with more than 4000 data samples of each region. In order to verify the predictive ability of this method, this paper uses the data samples from June 1 to November 20 as the model training sample set, and 24 data samples from the data in November 21 as the test set. The changes have been marked in red in the paper.

• We have improved the whole process of training the prediction model. The modifications are marked in red in the paper.The whole training process is as follows: firstly, the FL server sends down the initialization parameters of the prediction model to each local server, and then the global model parameters are sent down in subsequent rounds. The local servers use the initialization parameters for the first training, and each subsequent training is performed locally with the global parameters received by the FL server to obtain the local model parameters. In each round of communication, each operator uploads its respective local model parameters to the federation server for a new round of model aggregation and update until the model performance requirement or the specified number of communication rounds is reached.

The main innovation of this paper is to propose the FL-LSTM-BPNN model for the first time and apply the Federated Learning training approach in the training of PV power prediction models. At present, PV power prediction still suffers from the problem of not having sufficient number and feature variety of datasets and data privacy leading to data silos in practical applications. In contrast, the application of Federated Learning achieves high accuracy prediction of PV power while protecting data privacy. In the laboratory of this paper, it is demonstrated that models that can predict the power generation in other regions can also be trained with data from some regions with good prediction results.

Round 2

Reviewer 1 Report

Thanks, the authors answered all my questions.

Reviewer 3 Report

The paper can be published