Model Predictive Control of Humidity Deficit and Temperature in Winter Greenhouses: Subspace Weather-Based Modelling and Sampling Period Effects

Round 1

Reviewer 1 Report

The study developed a model predictive control (MPC) controller for greenhouse humidity deficit and temperature control. The study is of interest in the area and fits the journal’s scope. The main limitation of the study is the novelty, which is not clear.

(1)    There are many studies on using MPC for greenhouse control. Please conduct a literature review of the relevant studies

(2)    What is the novelty of this study? Please discuss the research gaps and the new contributions of this paper to the literature.

(3)    Section 2.1: what is the device for humidity control, humidifier?

(4)    What is the physical meaning of each term in the objective function of MPC? Please explain the control objectives in detail.

(5)    The prediction horizon of 16 minutes is very short considering the variation of weather conditions. In addition, the response of greenhouse temperature should be long than that. Please justify.

None

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper titled "Model Predictive Control of Humidity Deficit & Temperature 2 in Winter Greenhouses: Subspace Weather-Based Modelling & 3 Sampling Period Effects" proposes a model that considers the effects of climate on temperature control in greenhouses during the winter. The article is interesting and shows novelty. Some comments to the authors:

In Figure 2, when presenting all the information for each of the days, it is very loaded, so I would recommend better presenting the days or weeks that are typical during the data collection period.

The quality of Figure 3 needs to be improved. In the case of Figure 4, they could increase the size to improve it.

When working with time series, it is expected that the days used for training and/or evaluation will be consecutive. Why use the days 1-3, 31-38 instead of days 28-38?

To improve Figure 5, it would be better to reduce the content of the legends and place the most relevant information in the previous paragraph.

Improve the quality of Figure 7 and Figure 9.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper demonstrates the application of system identification (subspace) on greenhouse data and consequently the application of MPC using the identified model. I miss novelty.

Remarks:

- It seems that the authors used the identified model in the MPC (optimization). Consequently, the optimized control signals are also applied to the identified model. In other words, there is no mismatch between model in the controller and model where the contorl signals are applied. This is not a practical assumption since there are always differences. Can the authors also say something about the robustness of their method, via simulations? What if there is a model mismatch?

- Why did the authors not apply the optimized control signals to a real greenhouse? Or did they?

- The references need a serious update. The field of optimal control/MPC in greenhouses started around 1990. Most references are after 2000 so around 10 years of research is skipped.

- In (3), u(t) is defined. This also contains signals that we cannot change (disturbances). However, in the MPC, these are the signals that are optimized? We normally have a u(t) and a d(t)? 

How do you know the future HD_o, T_o and LX_i without modelled uncertainty?

- In (6), the cost is defined. The authors should also write down the optimization problem that is solved. What are the constraints? 

- The subspace model, how many states does it have? Is there a delay time defined?

- Why P=16 minutes? and what is the argument for the Q and R values? There is a phrase stating that the controllability is worsened. However, a linear model is controllable or not. This cannot become better or worse.

- In practice, what is a standard way to choose the sample period? Why deviate from that and test a couple of sample periods?

- Why is MPC applied here? The problem seems rather simple from a control point of view. Why not a PID, for example? 

- The last phrase of the conclusion is not clear to me. What is a more practical MPC?

Please read again carefully the document. I detected quite some grammar mistakes and phrases that do not make sense (grammar wise).

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The revised version can be accepted. 

Author Response

Thank you very much for taking time out of your busy schedule to consider this article.

Reviewer 2 Report

I am satisfied with the answers provided by the authors.

Author Response

Thank you very much for taking the time to review this paper despite your busy schedule.

Reviewer 3 Report

I have the following questions/remarks left:

- What is the difference between reference [27] and the work presented in this paper?

- What are the initial state variables of the model used in the MPC? The states do not represent any physical variables (subspace sysid) so how to initialize these? Please write down the optimization problem that is solved in the MPC.

- the window opening is used to control the indoor humidity and temperature. It seems to me that the wind speed/direction is having a significant influence on this transfer? Why is this not considered?

- the authors try to give a reasoning for choosing the sample period. However, in practice, the opening of the window has not this sample period. This is more done on an hourly basis. Is it possible to test the controller at this sample rate? It might be that the authors can show that the sample period needs to be set smaller with respect to current practice for better performance, that would be intresting.

- The authors write:

Simulation experiments using different models have been discussed in references [27, 481 28]. In this system, it has been verified that the degradation of the control performance owing to errors between the model in the controller and the applied model is marginal and can be compensated for by the sequential correction operation of the MPC.

   This is not a reason to assume that the same holds here. Or is the work in       the references very similar? Every application and model is different and         these kind of generalisations seem tricky.

- although the presentation of the results is improved, the contribution of this paper is still not so clear to me. The authors fit subspace models on different ranges of the data and then use these models in a MPC and apply the control signals on the same identified models. It is obvious that the MPC works. What can we actually learn from the results?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 3

Reviewer 3 Report

I would like to thank the authors for their answers. I agree with accepting the manuscript although I still think that the novelty (to any field) is minor.