Meteorological Variables Forecasting System Using Machine Learning and Open-Source Software
and David Raimundo RivasRound 1
Reviewer 1 Report
Paper is interesting but I have found in the paper so much novelty is described properly authors should explain and see the attachment for suggestions to paper improvement. I have accepted the paper with major revision.
Comments for author File: 
Comments.pdf
Author Response
Dear reviewer, Thanks for your comments. The answers are in the attached file.Author Response File: Author Response.pdf
Reviewer 2 Report
The paper presents a prediction of Meteorological variables e.g. temperature, humidity etc..
The paper lacks any significant contribution. It is a well-established fact that the prediction of Meteorological data needs physical models. many physical models even outperform the ML models. The presented work can be considered as tutorial or undergrad assignment at best.
It also lacks the proper understanding of ML modeling and evaluation metrics.
Author Response
Dear reviewer, Thanks for your comments. The answers are in the attached file.Author Response File: Author Response.pdf
Reviewer 3 Report
This is a useful study on using ML techniques to predict near surface weather variables
A very relevant recent study, similar to yours, but using a global data set ishttps://journals.ametsoc.org/view/journals/mwre/aop/MWR-D-22-0107.1/MWR-D-22-0107.1.xml
Overall the study is interesting and reasonable well done. However, in addition to the specific points listed below several important points should be addressed
- You have to better explain why you want to predict these, e.g. wind is generally independent of other variables used here (see article above), so why would you predict wind as function of T, solar radiation ? (and not using other data like reanalysis, weather forecast wind speed at 850 hPa and orographic height etc)- one can also see from your fit that wind is the most difficult
-the "verification" of predictions using just two days gives a reasonable illustration but generally is not sufficient. Couldn't you use more, eg every 2nd day of year for training and verification or selected random days?
- do you really need 5 min output for (energy applications)? or could one pre filter data?
-please better explain and summarise in Conclusions the apparent noise/bias introduced by some of the methods (decision tree/NN) and if this noise can easily be filtered etc
-towards the end of the manuscript there are several errors in the Tables (see comments), please carefully correct and rewrite Conclusions, also referring to the above points
-
please read, it should help to further improve your manuscript and the meteorological background, ie observation, forecast model, reanalysis data
-l14:"..for meteorological variables using different Machine Learning techniques implemented in a free software (Python)" -> "... for meteorological variables have been performed using different Machine Learning techniques as part of Python open source software.
-l15:"are the follower: multiple .." -> "include multiple ..."
-l34-36, rewrite .. maybe sth like" The 4-dimensional state of the atmosphere, the actual weather, is characterised by the wind, temperature and humidity variables forced by radiative fluxes and surface latent and sensible heat fluxes. The local climate denotes usually the mean state of the atmosphere over a 20-30 year period for a given location and day (or season) of the year.
-l47 "prediction models" which ones? Note that weather predictions already exist that are skilful up to 10 days and extended range predictions up to 4 weeks that can predict anomalies. You should distinguish these from the applications you cite afterwards
-l58: "having prediction system is costly" well yes but it can give you the 4d state of the atmosphere while current AI only does a set of selected 2D variables which is of course cheaper
-l100-104: be more precise, presumably you use 2m temperature and 10 m wind speed , 2m relative humidity (dew point)? and solar radiation (but which one incoming solar radiation SW+LW? or only SW? or net solar radiation at the surface)
-l121: unfortunately 2 days to validate is very short, you can be ok by chance
-l143: what about covariances and terms b1,2, ie XiXj ?
-l281-283: please check the definitions, for rMSE you say "prediction error in the analysis time", what do you mean by analysis time, while for R^2 you say in the process of running the prediction - However a usual definition of R^2 is the explained variance, is this what you mean by "efficiency"? please explain better following statistics definition
-Figure 5: It is obvious from Figure 5 that the decision tree and random forest add noise to the the data and therefore have lower R^2. Should the input dat be pre filtered or does the decision tree require and additional low pass filter? , we do not need 5 min T variations output for renewable energy, 30-60 min is enough. On the other side there seems to be a bias issue in the NN.
-Table 8: why do you suddenly have units of W/m2 for relative humidity? It looks like there is a mistake and you have duplicated Table 9 there. Correct!
-Table 9: it is in Spanish!!
-Conclusions not very informative, rewrite!
Author Response
Dear reviewer,
Thanks for your comments. The answers are in the attached file.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
I have accepted the article for publication in journal, authors respond to all reviewers comments.
Author Response
We thank for your fast answer, and constructive review and careful reading as it helps us to improve the update manuscript.
Reviewer 2 Report
Accept
Author Response
We thank for your fast answer, and constructive review and careful reading as it helps us to improve the update manuscript.
