Predictive Modeling of Photovoltaic Panel Power Production through On-Site Environmental and Electrical Measurements Using Artificial Neural Networks

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The topic is interesting but the article lacks many important details. It can be evaluated after careful and major revision according to the following comments.

1.      ANN and its implementation are well known and accessible in literature. The given details are unnecessary.

2.      This topic is widely studied by researchers using state-of-the-art machine/deep learning techniques/algorithms. Why the authors just used ANN? And what is the improvement?

3.      The developed models should be compared with other well and widely studied algorithms.

4.      Related literature is not well reviewed. What is the improvement?

5.      It is written that “………………collected data was compared with measurements from a commercial meteorological station….”. The comparison is not well discussed. And how the difference is dealt? The normalization is according to one dataset, as mentioned later.  

6.      How voltage and current measurements are taken in designed prototype?

7.      There are several devices available. The design is not clear? What is different?

8.      How the models are trained? The specific implementation and architectures are not discussed. Explain in a well manner.

9.      The dataset is not clear. Provide details of dataset and train-test split.

10.  What is the purpose of figure 12? It is not mentioned in text body.

11.  Multiple Linear Regressor Model development is not discussed. It is just mentioned. The comparison should be with other well established and studied models.

12.  The title is not suitable and confusing. It can be revised.

13.  The use of forecasting model is not clear? Which non-autoregressive algorithm is used? How?

 

Comments on the Quality of English Language

Nil

Author Response

Dear reviewers,

We express our gratitude for the valuable feedback provided on the submitted manuscript. Each reviewer's comment has been respectfully addressed, accompanied by a comprehensive explanation. We appreciate the time and effort invested in reviewing our work and hope our responses adequately address any concerns raised.

## Inquiry 1
ANN and its implementation are well known and accessible in literature. The given details are unnecessary.

## Answer 1:

The authors appreciate the reviewer's perspective on the familiarity and accessibility of Artificial Neural Networks (ANN) and their implementation in the existing literature. ANNs have indeed been extensively studied and documented. However, it is essential to consider the context in which the details are provided and the Journal's primary focus. In this case, additional details may be necessary to comprehensively understand the use of ANNs as universal regressors and the steps followed during methodology. Effective communication often involves striking a balance between providing necessary information and avoiding unnecessary redundancy.

## Inquiry 2
This topic is widely studied by researchers using state-of-the-art machine/deep learning techniques/algorithms. Why the authors just used ANN? And what is the improvement?

## Answer 2:
The reviewer's inquiry regarding using Artificial Neural Networks (ANNs) in this study is valid, given the prevalence of state-of-the-art machine and deep learning techniques and algorithms in current research. It is important to note that the authors have opted for ANNs for specific reasons:

- Relevance of ANNs: The introduction and state-of-the-art shows the relevance and suitability of ANNs in the addressed application due to their ability to model complex(non-linear) relationships in spatial data.
- Technical Gap: The identified technical gap involving the use of measurements in the proximity of the panel's location could improve the ANN performance by minimizing the RMSE of power estimations. Then, the on-site measurements requirement is covered by developing an IoT device for continuous logging (every 5 minutes) in exposed-weather conditions. 
- Implementation Considerations: ANNs can indeed be implemented efficiently on microcontrollers, which might align with the authors' intentions for further practical applications, especially with limited computational resources.

Then, to better understand the improvement obtained by the use of ANNs in this specific context, a particular examination of the methodology, experimental results, and comparative analysis with other topologies were offered. Additionally, the paper's state-of-art has been modified to highlight why the authors selected ANN and make clear the research's main contribution. 

## Inquiry 3.
The developed models should be compared with other well and widely studied algorithms.

## Answer 3:
The recommendation to compare the developed models with other well-established and widely studied algorithms is valid. However, it is necessary to consider the specific context outlined in the present research:

1. Elaborateness of ANN Development: The development of ANNs involved a considerable effort to explore various topologies(models): number of inputs, number of hidden layers/nodes, selected activation function, learning rate, momentum, and training cycles. While the paper reports only a limited number of topologies, it is important to consider that an extensive effort was expended in training at least 35 different topologies. The training work indicates a rigorous exploration of neural network architectures. Thus, more model results are added to support the topologies exploration.
2. Primary Focus on On-site Measurements: The study's central focus is examining on-site measurements' impact on the model. Given this emphasis, the paper should prioritize a detailed analysis of the ANN models, as they are specifically designed to address this aspect.
3. Forecasting Potential: The versatility of the ANN topology, with the potential for use in forecasting with minor modifications, adds another matter of relevance to the study's primary objectives.
4. Measurement Evaluation Efforts: The action expended on evaluating measurements, likely contributing to the study's findings' overall quality and reliability.

In addressing the suggestion to compare with other algorithms, the authors prefer to discuss why ANN topologies were chosen over alternatives, how they performed compared to other topologies, and how this choice aligns with the specific objectives and challenges of the study. Furthermore, the reviewer should consider that there are limitations in the comparison due to the unique focus of the research.

## Inquiry 4
Related literature is not well reviewed. What is the improvement?

## Answer 4
The comment regarding the review of related literature needing to be more comprehensive is valid and prompts consideration of potential improvements. The following points may serve as an answer that will be included in the paper:

1. Utilization of Online Data: Indeed, a significant portion of the existing literature relies predominantly on online data sources for meteorological stations. To address this limitation, the authors' contribution relies on gathering accurate on-site data directly measured in the PV panel.
2. Emphasis on Representative Data: Numerous prior works acknowledge the significance of constructing models with representative data. This study's improvement could lie in explicitly detailing the efforts made to ensure data representativeness, potentially discussing the selection criteria for on-site measurements and their relevance to real-world conditions.

3. Custom Measurement Device: The development of a custom device for measuring and modeling PV panels is a noteworthy innovation. This improvement could be highlighted by explaining the device's design, capabilities, and how it contributes to the accuracy and reliability of the data collected.

4. Quantitative Emphasis on On-site Measurements: The emphasis on the quantity of on-site measurements is another distinguishing feature. To address this improvement, the research can elaborate on the advantages of a large quantity of on-site measurements, such as improved statistical robustness, trend identification, and model generalization.

In essence, the improvement lies in the research's meticulous approach to data collection and modeling, which differentiates it from prior literature that primarily relies on online data sources and does not delve as deeply into the nuances of on-site measurements and their potential impact on model accuracy and applicability.

## Inquiry 5
It is written that "………………collected data was compared with measurements from a commercial meteorological station….". The comparison is not well discussed. And how the difference is dealt? The normalization is according to one dataset, as mentioned later.
## Answer 5
This section has been revised and expanded to provide a clearer explanation.

## Inquiry 6
How voltage and current measurements are taken in designed prototype?

## Answer 6
The prototype measures all environmental variables and voltage-current generated by the panel. Thus, every 5 minutes, the prototype logs time, temperature, luxes, humidity, and the power (Voltage and Current) generated by the PV. This creates an instance to train the ANN regressor.
Additionally, the section has been revised and expanded to provide a clearer explanation.

## Inquiry 7
There are several devices available. The design is not clear? What is different?

## Answer 7
We have updated and expanded the section to offer a more concise and understandable explanation.

## Inquiry 8
How the models are trained? The specific implementation and architectures are not discussed. Explain in a well manner.

## Answer 8
We have updated and expanded the section to offer a more concise and understandable explanation.
## Inquiry 9
The dataset is not clear. Provide details of dataset and train-test split.

## Answer 9
We have updated and expanded the section to offer a more concise and understandable explanation.

## Inquiry 10
What is the purpose of figure 12? It is not mentioned in text body.
## Answer 10
There was a mistake in the figure's cross-references.

## Inquiry 11
Multiple Linear Regressor Model development is not discussed. It is just mentioned. The comparison should be with other well established and studied models.
## Answer 11
We have updated and expanded the section to offer a more concise and understandable explanation.

## Inquiry 12
The title is not suitable and confusing. It can be revised.
## Answer 12

The title has been revised and the new proposal is: *Estimation of Photovoltaic Power Output using Onsite Measurements Gathered Through an IoT Logger Utilizing an Artificial Neural Network Model.*

## Inquiry 13
The use of forecasting model is not clear? Which non-autoregressive algorithm is used? How?

## Anser 13
We have updated and expanded the section to offer a more concise and understandable explanation.

Reviewer 2 Report

Comments and Suggestions for Authors

The paper presents an interesting mathematical study using ANN topologies for forecasting the output power of solar modules. However, pratical and scientific relevance for the photovoltaic community are very limited, due to the following reasons:

1. Training and validation of the ANN is only done with clear sky conditions, and hence, the relevance for other weather conditions is not given.

2. The model results in an increase of output power with temperature, which is a completely wrong conclusion in the physical world.

3. By chance, during clear sky conditions, output power increases with increasing humidity and temperature, in the real physical world this does not make any sense. I assume that for real weather conditions, RMSE will be much higher!

4. Traning and validation is only done with a very small solar module not described further (probably, a thin film module, which is not very representative, as most solar modules are crystalline Si)

You say: "we compared them to those of a commercial station at the Technological Institute of Morelia, Mexico." The results are, however, not shown and cannot be included into your conclusion.

I recommend to do a sound sensitivity analysis and to combine your ANN study with some physical implications. Your study must be validated by different kind of weather condistions: not only clear sky and by different type of solar modules (especially c-Si, not only the small module used in the study).

Comments on the Quality of English Language

second person (we conclude...) should be avoided

Author Response

Dear reviewers, 

We express our gratitude for the valuable feedback provided on the submitted manuscript. Each reviewer's comment has been respectfully addressed, accompanied by a comprehensive explanation. We appreciate the time and effort invested in reviewing our work and hope that our responses adequately address any concerns raised.

 

## Inquiry 1
Training and validation of the ANN is only done with clear sky conditions, and hence, the relevance for other weather conditions is not given.

## Answer 1
A piece of very short information about the weather in Mexico is available in this dataset, even in nighttime, rainy, and cloudy conditions. However, after carefully analyzing correlation plots, it has been observed that low or absent lighting conditions are associated with low or non-power values. Thus, in this first paper, the authors prefer to show Mexico's average weather. Nevertheless, the dataset remains growing with new measurements over the days for further research.

## Inquiry 2
The model results in an increase of output power with temperature, which is a completely wrong conclusion in the physical world.

## Answer 2:
It is well known that temperature affects the general efficiency of semiconductors, the main component of solar panels. Nevertheless, solar panels' common performance is their exposition to solar radiation. Therefore, the charts indicate the correlation between all environmental variables happening simultaneously: The solar panel's performance may be affected by high temperatures resulting from high solar radiation (luxes). To capture the negative effects on the generated power, the model measures the temperature of the solar panel. This point has been included in the paper to provide clarity.

## Inquiry 3:
By chance, during clear sky conditions, output power increases with increasing humidity and temperature, in the real physical world this does not make any sense. I assume that for real weather conditions, RMSE will be much higher!

## Answer 3:
The charts and correlation plots show that the humidity is high under low light and temperature levels; in the night and early morning periods. Additionally, the plots show that humidity is low when output power is high.

## Inquiry 4:
Traning and validation is only done with a very small solar module not described further (probably, a thin film module, which is not very representative, as most solar modules are crystalline Si)
## Answer 4:
The paper aims to avoid generalizing the model for solar panel modules or sizes. Instead, the paper pretends to show the general procedure to use and develop ANN as regressors supported by a measurement IoT device.

## Inquiry 5:
You say: "we compared them to those of a commercial station at the Technological Institute of Morelia, Mexico." The results are, however, not shown and cannot be included into your conclusion.

## Answer 5
Figure 5 has been modified to clarify the data comparison between both devices.

## Inquiry 6:
I recommend to do a sound sensitivity analysis and to combine your ANN study with some physical implications. Your study must be validated by different kind of weather condistions: not only clear sky and by different type of solar modules (especially c-Si, not only the small module used in the study).

## Answer 6
This first approach can be extended to larger systems with different types of solar panels in the future with larger and complex systems.

Reviewer 3 Report

Comments and Suggestions for Authors

Interesting paper. A few comments:

The complexity of the developed method is quite high. Wasn’t it easier to use neural network(s)?

Line 123: “All are function” => “All are functions”

Line 127: “The maximum admissible magnitude values in voltage, “ => “The maximum admissible magnitude values of voltage, “

Equation (9) seems to be out of the context.

Line 329: “The bloc unit time” => “The block unit time” (probably).

Line 349: “The main issue remains its implementation” => “The main issue Romains is its implementation”

Line 356-357: “is to minimizing” => “is to minimize”.

Author Response

Dear reviewers, 

We express our gratitude for the valuable feedback provided on the submitted manuscript. Each reviewer's comment has been respectfully addressed, accompanied by a comprehensive explanation. We appreciate the time and effort invested in reviewing our work and hope that our responses adequately address any concerns raised.

 

Answer:

All the comments have been reviewed and changes are considered in the newest manuscript version. Also, we have updated and expanded diverse sections to offer a more concise and understandable explanation.

Regarding equation number nine, the paper does not include that number of equations.
 

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Nil

Comments on the Quality of English Language

Nil

Author Response

Dear reviewer,

Thank you for reviewing our manuscript and for providing valuable feedback.

We are pleased to inform you that we have carefully considered your comments and suggestions and have made the necessary updates to the paper.

We kindly request you to review the updated version of our paper. We believe the modifications will meet your approval and further the paper's contribution to the field.

Again, Thank you for your constructive feedback and continued attention to our work.

Reviewer 2 Report

Comments and Suggestions for Authors

Dear authors, thank you for your answers:

## Answer 1
A piece of very short information about the weather in Mexico is available in this dataset, even in nighttime, rainy, and cloudy conditions. However, after carefully analyzing correlation plots, it has been observed that low or absent lighting conditions are associated with low or non-power values. Thus, in this first paper, the authors prefer to show Mexico's average weather. Nevertheless, the dataset remains growing with new measurements over the days for further research.

## my reply 1:

In this case, you should mention in your chapter "discussion and concludions", that your model had only been trained and validated under clear sky conditions and that a training and validation under cloudy/overcast sky conditions is subject to further studies.

## Answer 2:
It is well known that temperature affects the general efficiency of semiconductors, the main component of solar panels. Nevertheless, solar panels' common performance is their exposition to solar radiation. Therefore, the charts indicate the correlation between all environmental variables happening simultaneously: The solar panel's performance may be affected by high temperatures resulting from high solar radiation (luxes). To capture the negative effects on the generated power, the model measures the temperature of the solar panel. This point has been included in the paper to provide clarity.

## my reply 2:

I am not refering to the difference between module and ambient temperature, I mean that PV output power correlates with a negative temperature coefficient with temperature (assuming constant irradiation). This point should be discussed in results and discussion. Have you trained and analyzed your modeling with JUST irradiance data? That would be interesting.

## Answer 3:
The charts and correlation plots show that the humidity is high under low light and temperature levels; in the night and early morning periods. Additionally, the plots show that humidity is low when output power is high.

## my reply 3:

this is why, the estimator based on temperature and humidity variables (line 326) gives the highest error with 1.52. this high error should be discussed in your last chapter "discussion", I think that this estimator based on temperature and humidity does not make sense due to the facts explained above. This is also supported by Fig. 10 a-d, this sould be mentioned in your last chapter "discussion"

## Answer 4:
The paper aims to avoid generalizing the model for solar panel modules or sizes. Instead, the paper pretends to show the general procedure to use and develop ANN as regressors supported by a measurement IoT device.

## my reply to 4:

o.k., you should make it really clear in the conclusion, as well.

## Answer 5
Figure 5 has been modified to clarify the data comparison between both devices.

## my reply to 5:

o.k., thank you

## Answer 6
This first approach can be extended to larger systems with different types of solar panels in the future with larger and complex systems.

## my reply to 6:

o.k., you should mention exactly this in "discussion and conclusion"

 

Author Response

Dear reviewer,

Thank you for reviewing our manuscript and for providing valuable feedback.

We are pleased to inform you that we have carefully considered your comments and suggestions and have made the necessary updates to the paper.

We kindly request you to review the updated version of our paper. We believe the modifications will meet your approval and further the paper's contribution to the field.

Again, Thank you for your constructive feedback and continued attention to our work.

 

# Inquiry 1

In this case, you should mention in your chapter "discussion and concludions", that your model had only been trained and validated under clear sky conditions and that a training and validation under cloudy/overcast sky conditions is subject to further studies.

 

## Answer 1
The section has been revised and expanded to provide a clearer explanation.

## Inquiry 2

I am not refering to the difference between module and ambient temperature, I mean that PV output power correlates with a negative temperature coefficient with temperature (assuming constant irradiation). This point should be discussed in results and discussion. Have you trained and analyzed your modeling with JUST irradiance data? That would be interesting.

## Answer 2

Dear reviewer, thanks for the insightful feedback. The authors understand the concern regarding the negative temperature coefficient's impact on PV output power, assuming constant irradiation. As previously answered, we primarily focused on showing the correlation plots between the gathered data in the results section. Numerical correlation tests give positive values for power vs. lux and power vs. Temperature; we are uploading screenshots of the correlation matrix and bidimensional scatterplots for a better understanding.

Regarding the inquiry about training and analyzing the model with only irradiance data, one model was trained primarily with irradiance data. However, its performance was not good enough compared with models with two or more additional parameters.

## Inquiry 3
this is why, the estimator based on temperature and humidity variables (line 326) gives the highest error with 1.52. this high error should be discussed in your last chapter "discussion", I think that this estimator based on temperature and humidity does not make sense due to the facts explained above. This is also supported by Fig. 10 a-d, this sould be mentioned in your last chapter "discussion"

## Answer 3
The section has been revised and expanded to provide a clearer explanation.

## Inquiry 4
o.k., you should make it really clear in the conclusion, as well.

## Answer 4

The section has been revised and expanded to include this explanation.

## Inquiry 5
NA

## Answer 5
NA

## Inquiry 6

o.k., you should mention exactly this in "discussion and conclusion"

 

## Answer 6

The section has been revised and expanded to include this explanation.

 

Author Response File: Author Response.pdf