Evaluating Time-Series Prediction of Temperature, Relative Humidity, and CO₂ in the Greenhouse with Transformer-Based and RNN-Based Models

Round 1

Reviewer 1 Report

The objective of this study was to predict the changes in the temperature, relative humidity (RH), and CO2 concentration in a greenhouse that would occur after 1 and 3 h, using four external climate data, three internal data, and six actuator data. For this purpose, a transformer-based model (Autoformer) and a simple linear model (DLinear) were employed, and their prediction performance was evaluated. Overall, DLinear consistently outperformed Autoformer. The results indicated that the transformer-based model, Autoformer, was not as effective in predicting greenhouse environments as the simple linear-based model, DLinear. While DLinear’s performance was generally satisfactory, it did not perform as well in predicting CO2 concentrations.

This is an interesting study, but there are still some issues that need to be addressed before publication.

1 To verify the generalization of the model, authors should use more datasets.

2 Authors should compare more recently published methods as baseline methods.

3 Authors should discuss more time series related work in the introduction, such as [1-3].

[1] Refined nonuniform embedding for coupling detection in multivariate time series[J]. Physical Review E, 2020, 101(6): 062113.
[2] SleepPrintNet: A multivariate multimodal neural network based on physiological time-series for automatic sleep staging[J]. IEEE Transactions on Artificial Intelligence, 2020, 1(3): 248-257.

Minor editing of English language required

Author Response

Dear Reviewer(s),

First, we would like to thank you for your critical and constructive feedback on this manuscript. The comments and suggestions were very thorough and useful in improving the quality of the paper. We have taken them fully into account in revision, and we strongly believe that all inputs have significantly increased the scientific value of the revised manuscript.

The manuscript has been substantially revised. In the original version, we compared the performance of only two models (Autoformer and DLinear). However, in response to the reviewer’s recommendations to include additional SOTA (state-of-the-art) models, we conducted further experiments with two newer SOTA models (LSTM and SegRNN) and have now fully integrated the results of the experiments into the manuscript. This comprehensive revision process took more time than expected, especially during the year-end & new-year holiday season, which resulted in the delay of the resubmission. Nevertheless, we appreciate the input as it has helped us to improve the manuscript. We have thoroughly reviewed all the comments and have made significant improvements to the manuscript accordingly.

We are submitting the revised manuscript as a separate Word file alongside this rebuttal. In the revised manuscript, any newly added or modified sentences are highlighted in red. Below, we attached a separate Word file with detailed response to your comments. 

Thank you very much.

Author Response File: Author Response.pdf

Reviewer 2 Report

Dear Author

The topic of the manuscript is a current a will be interesting the possible aplications in prediction.

Some general comments

Title: The title is ambiguos, because transformer based model is general (at the end directly model was better), may be you can be more specific, prediction of climate variables in greenhouse by using ...

abstract: Can you complement with specific comments in the manuscript

Introduction: In opinion of this reviewer, authors must (here or in merthods) references about why or how the fdata adquisition impact over the prediction. Of course, your objective is the prediction of tree environment variables, however, the objective is quality prediction in order to have an useful tool. That is why, the quality of data is basic.

M&M; As a consequence in this section, could be necessary to test or explain, the source of the statistical environmental variables. Authors mentioned just in a superficial way.

For instance, explain why 1 hour or 3 hour after, is the same time for Temperature (°C) Humidity (%) and CO2 (ppm), inside/outside?

Why are using these statistical indicators?

On the other hand, when you mentioned the actuadors, it is convenient to describe the brand or precision of sensors, range etc.

YES, I understand your objective is the prediction, but, another one, it is necessary prediction with quality, and the good prediction depend of the quality of data source.

Results and discussion. As you well mentioned, the results are not very good, and in the discussion some of your afirmations are in general (in the manuscript). Most of the numerical tools pretend the prediction or description of a part of reality, but authors must be specific, what was the results of their specific research, what was the diference with others and what is the main aportation of your results.

Other specific comments are in the manuscript

Comments for author File: Comments.pdf

Author Response

Dear Reviewer(s),

First, we would like to thank you for your critical and constructive feedback on this manuscript. The comments and suggestions were very thorough and useful in improving the quality of the paper. We have taken them fully into account in revision, and we strongly believe that all inputs have significantly increased the scientific value of the revised manuscript.

The manuscript has been substantially revised. In the original version, we compared the performance of only two models (Autoformer and DLinear). However, in response to the reviewer’s recommendations to include additional SOTA (state-of-the-art) models, we conducted further experiments with two newer SOTA models (LSTM and SegRNN) and have now fully integrated the results of the experiments into the manuscript. This comprehensive revision process took more time than expected, especially during the year-end & new-year holiday season, which resulted in the delay of the resubmission. Nevertheless, we appreciate the input as it has helped us to improve the manuscript. We have thoroughly reviewed all the comments and have made significant improvements to the manuscript accordingly.

We are submitting the revised manuscript as a separate Word file alongside this rebuttal. In the revised manuscript, any newly added or modified sentences are highlighted in red. Below, we attached a separate Word file with detailed response to your comments. 

Thank you very much.

Author Response File: Author Response.pdf

Reviewer 3 Report

Notes to authors:

1. The period for carrying out the observations is very short and does not take into account the whole variety of fluctuations of the meteorological conditions outside the greenhouse from year to year;

2. Taking into account the previous note, the statistical evaluations made in view of the applicability of the two Autoformer and DLinear models are imprecise and in some sense misleading;

3. The conclusions are very general and insufficiently informative due to an incomplete set of real data. In this case, the number of data for T, RH and CO2 is not so important, but the number of different combinations of these parameters outside the greenhouse.

4. The regularities that impressed the authors are true, but they are known to Agrometeorology and Agronomy as sciences examining different environments and plant ecosystems and their interaction. A similar kind of research was done in Europe already in the middle of the 20th century.

It is necessary to make minor corrections to the English text in order to improve the style of expression and to clarify the terminology.

Author Response

Dear Reviewer(s),

First, we would like to thank you for your critical and constructive feedback on this manuscript. The comments and suggestions were very thorough and useful in improving the quality of the paper. We have taken them fully into account in revision, and we strongly believe that all inputs have significantly increased the scientific value of the revised manuscript.

The manuscript has been substantially revised. In the original version, we compared the performance of only two models (Autoformer and DLinear). However, in response to the reviewer’s recommendations to include additional SOTA (state-of-the-art) models, we conducted further experiments with two newer SOTA models (LSTM and SegRNN) and have now fully integrated the results of the experiments into the manuscript. This comprehensive revision process took more time than expected, especially during the year-end & new-year holiday season, which resulted in the delay of the resubmission. Nevertheless, we appreciate the input as it has helped us to improve the manuscript. We have thoroughly reviewed all the comments and have made significant improvements to the manuscript accordingly.

We are submitting the revised manuscript as a separate Word file alongside this rebuttal. In the revised manuscript, any newly added or modified sentences are highlighted in red. Below, we attached a separate Word file with detailed response to your comments. 

Thank you very much.

Author Response File: Author Response.pdf