the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 08 Mar 2022
Fortnight conditioning of historical data to improve short-term precipitation predictions
Abstract. The effects of changes in weather variables, including precipitation dependence on the days-of-the-week, have known applications in weather predictions. However, the use of these effects to improve weather forecasting has not been determined. Here we investigate if conditioning past data somehow by considering the days-of-the-week helps us to obtain the better short-term time series prediction for precipitation. Especially, we demonstrate that short-term time series prediction of precipitation up to 2 h ahead can be improved using the data points of the days whose differences from the current day are multiples of 14. For short-term predictions, we employ infinite-dimensional delay coordinates (Hirata et al., Sci. Rep. 5, 15736, 2015) to reconstruct the underlying dynamics. Although the results demonstrate that the two-week periodicity seems to exist in the weather at Tokyo, and thus some anthropogenic activities could influence weather, the mechanism of the influence remains unclear.
- Preprint
(138 KB) - Metadata XML
-
Supplement
(4 KB) - BibTeX
- EndNote
Status: closed
-
CC1: 'Comment on npg-2022-9', Dmitri Kondrashov, 24 Mar 2022
I am simply not convinced by this paper, it is very short with one figure and is not up to the standards and depth expected for NPG. Authors need to heavily revise and extend the manuscript to improve presentation and their arguments. Hopefully my comments below are helpful.
The authors argue that short-term (2hr ahead) time series prediction for precipitation at Tokyo station in 1-min sampling can be improved by using data two weeks in the past and some form of analogs method. This is similar to looking for needle in a haystack and I find it very doubtful without additional analysis and presentation. First of all it would be helpful to show time series. Secondly, are there any periodicities in the time series itself by using classical spectral analysis methods? Finally, they should think on how to better present and illustrate their prediction method, perhaps using some toy model data, not simply as a short appendix.
Citation: https://doi.org/10.5194/npg-2022-9-CC1 -
AC1: 'Reply on CC1', Yoshito Hirata, 08 Apr 2022
Dear Dr. Dmitri Kondrashov,
Thank you very much for providing us your valuable comments.
Your comments helped us to explain, more deeply, the dataset of precipitation we have.
Thus, we would like to reply your comments as the attached PDF file.
We hope that we have covered all of your comments and can clarify the nature of the precipitation dataset.
Yours sincerely,
Yoshito Hirata
-
AC1: 'Reply on CC1', Yoshito Hirata, 08 Apr 2022
-
RC1: 'Comment on npg-2022-9', Anonymous Referee #1, 28 Mar 2022
Review of the manuscript ‘Fortnight conditioning of historical data to improve short-term precipitation predictions’ by Yoshito Hirata and Yoshinori Yamada
The present manuscript tries to attribute short-range precipitation predictability in the large Tokyo megalopolis to the indirect effect of aerosols produced by anthropogenic activities, through their influence on the production of precipitation nuclei and optic effects.
The manuscript is very short not giving enough details for the appropriate reproducibility of the results. Moreover, the methodology, and the arguments in the discussion are very dubious and even not falsifiable, which is fundamental requirement in any scientific theory. Moreover, there are severe methodological shortcomings, described below. Giving those reasons, the manuscript is judged not reaching enough standards to be published in NPG.
The present study should be preceded by experiments with a toy minimal model, reproducing convection and precipitation mechanisms triggered by aerosol nucleation. Then, predictability experiments should be run by imposing some weekly periodicity to aerosol emissions to simulate the periodic anthropogenic forcing and seek whether any phase synchronizing is observed in precipitation. The predictability study described in the manuscript, obtained with timeseries only is far unsatisfactory due to the existence of a vast number of noncontrolled factors, beyond aerosols. It is thus very difficult to produce a convincing quantifiable attribution of the very-short term precipitation predictability to the aerosol’s forcing.
The applied methodology is dubious and impacted by severe pitfalls such as:
- The method of analogues is too little described; for instance, the analogs metric is not clear. Is it based on precipitation only? If yes, the analog’s distance is too strict.
- It is not clear if analogs are sought in an independent period of the validation period.
- The details of the AR model are not described. Other benchmark stochastic models should be tested.
- By forecast rank, authors mean error, so authors should precise that.
The unique figure presented is not fully discussed. There are results which are not understandable neither discussed such as: the bump in rank around the forecast delay 60-70 minutes for D=1; the reason why the predictability is larger when analogous are sought with D=14 than D=7. Authors present a very speculative unproven reason for that: ‘there is a period doubling bifurcation in the precipitation and that a week periodicity, if it exists, could be unstable’.
Citation: https://doi.org/10.5194/npg-2022-9-RC1 -
AC2: 'Reply on RC1', Yoshito Hirata, 08 Apr 2022
Dear Referee #1,
Thank you very much for providing your insightful comments.
We would like to reply your comments as the attached PDF file.
Especially, please let us know the names of some benchmark stochastic models if we should addn them
because we have alreadly included the AR model, the persistence model and the mean prediction model,
which ,we believe, cover the standard benchmark stochastic models.
Yours sincerely,
Yoshito Hirata
-
RC2: 'Comment on npg-2022-9', Anonymous Referee #2, 23 May 2022
This paper illustrates how conditioning a nowcast precipitation prediction on the calendar day improves a precipitation score.
The poor English syntax and grammar make the manuscript very difficult to follow. The lack of clear and detailed explanations of what is done make the manuscript impossible to assess. For example, the annex is not called in the main text and it does not clarify anything on the procedure that is used by the authors.
From what I see, I do not see how the paper is relevant in NPG, as I do not see a real conceptual innovation (the only innovation appeared in a paper already published by the authors).
Therefore, my appraisal of the paper is based on a guess of what was done to obtain the results.
Major points
The authors use time series with a time increment of one minute. Therefore, not only there is a seasonal cycle, but the time series also contain a diurnal cycle. If there is any cycle in the data, a Fourier transform should be able to detect it. The results reported in Figure 1 do not suggest any type of periodicity.
The methods section does not state how precipitation is predicted (e.g. what model?). Even the AR prediction is not clear. How are the authors certain that they do not over fit the data?
When they use the term “improve” (e.g. in the title), they should state with respect to what? The improvement over operational nowcasting from meteorological institutions should be demonstrated.
I feel that the reported result (better 2h forecast when taking D=14 day prior information) is only valid for the statistical scheme alluded to by the authors. Nothing proves or even suggests that this would hold in a “regular” nowcasting meteorological forecast.
The main result of the paper is based on Figure 1. But this figure does not prove anything, in particular for D=14. The authors have not tried other values of D, in particular larger values. The seasonal dependence is not discussed or even assessed. Why is there a “bump” for D=1? Precipitation differences of 0.006 mm (maximum value of the vertical axis in Figure 1) are not measurable by meteorological instruments. Therefore, the apparent minimum for D=14 cannot be measured in practice. This minimum of mean absolute error might not even be statistically significant (and it is obviously not physically relevant).
The right way to assess forecast schemes is to use cross validation procedures, i.e. at least by considering a training period and a separate validation period. Therefore, it is not even clear that the reported result is actually true.
Specific comments
Abstract: is precipitation dependence a weather variable? (or what is weather variable, and how do the authors define “precipitation dependence”?).
l. 15: Why and how the uni-modal relation (whatever that means) of aerosols and convective energy (why convective energy) is connected to the scattering and absorption of solar radiation?
I think that the authors miss the main point of predicting precipitation, as they treat zero values in the same way as non zero values.
The methods section is inappropriately unclear, especially for a journal like NPG. The first paragraph of section 3 should be in the methods section. The AR model is not defined properly. An order of 120 sounds like overfitting. Precipitation is not Gaussian, especially at minute time scales. An AR model is an obvious bad choice.
Why don’t the authors consider the hour of the day when they condition the forecast? They might avoid an aliasing phenomenon that could explain a fortnight conditioning.
The first paragraph of section 4 is incomprehensible, and is not related to analyses of the paper.
The Appendix section is not really informative on what is done in the forecast.
Conclusion
I cannot recommend the publication of this manuscript in NPG.
Citation: https://doi.org/10.5194/npg-2022-9-RC2 -
AC3: 'Reply on RC2', Yoshito Hirata, 30 May 2022
Dear Referee #2,
Thank you very much for reading our manuscript critically and providing your valuable feedback.
We have examined all of your comments and are replying to you as attached.
Overall, we believe that your valuable feedback helped us to strengthen our findings
because we could show that we can reproduce the similar results with 20-dimensional usual delay coordinates,
a gold standard in the field of nonlinear time series analysis.
Thus, it would be great if you read our response carefully and check that we have overcome all of your concerns.
Again, we deeply appreciate your critical comments, which helped us to demonstrate our findings in the different way.
Yours sincerely,
Yoshito Hirata
-
AC3: 'Reply on RC2', Yoshito Hirata, 30 May 2022
Status: closed
-
CC1: 'Comment on npg-2022-9', Dmitri Kondrashov, 24 Mar 2022
I am simply not convinced by this paper, it is very short with one figure and is not up to the standards and depth expected for NPG. Authors need to heavily revise and extend the manuscript to improve presentation and their arguments. Hopefully my comments below are helpful.
The authors argue that short-term (2hr ahead) time series prediction for precipitation at Tokyo station in 1-min sampling can be improved by using data two weeks in the past and some form of analogs method. This is similar to looking for needle in a haystack and I find it very doubtful without additional analysis and presentation. First of all it would be helpful to show time series. Secondly, are there any periodicities in the time series itself by using classical spectral analysis methods? Finally, they should think on how to better present and illustrate their prediction method, perhaps using some toy model data, not simply as a short appendix.
Citation: https://doi.org/10.5194/npg-2022-9-CC1 -
AC1: 'Reply on CC1', Yoshito Hirata, 08 Apr 2022
Dear Dr. Dmitri Kondrashov,
Thank you very much for providing us your valuable comments.
Your comments helped us to explain, more deeply, the dataset of precipitation we have.
Thus, we would like to reply your comments as the attached PDF file.
We hope that we have covered all of your comments and can clarify the nature of the precipitation dataset.
Yours sincerely,
Yoshito Hirata
-
AC1: 'Reply on CC1', Yoshito Hirata, 08 Apr 2022
-
RC1: 'Comment on npg-2022-9', Anonymous Referee #1, 28 Mar 2022
Review of the manuscript ‘Fortnight conditioning of historical data to improve short-term precipitation predictions’ by Yoshito Hirata and Yoshinori Yamada
The present manuscript tries to attribute short-range precipitation predictability in the large Tokyo megalopolis to the indirect effect of aerosols produced by anthropogenic activities, through their influence on the production of precipitation nuclei and optic effects.
The manuscript is very short not giving enough details for the appropriate reproducibility of the results. Moreover, the methodology, and the arguments in the discussion are very dubious and even not falsifiable, which is fundamental requirement in any scientific theory. Moreover, there are severe methodological shortcomings, described below. Giving those reasons, the manuscript is judged not reaching enough standards to be published in NPG.
The present study should be preceded by experiments with a toy minimal model, reproducing convection and precipitation mechanisms triggered by aerosol nucleation. Then, predictability experiments should be run by imposing some weekly periodicity to aerosol emissions to simulate the periodic anthropogenic forcing and seek whether any phase synchronizing is observed in precipitation. The predictability study described in the manuscript, obtained with timeseries only is far unsatisfactory due to the existence of a vast number of noncontrolled factors, beyond aerosols. It is thus very difficult to produce a convincing quantifiable attribution of the very-short term precipitation predictability to the aerosol’s forcing.
The applied methodology is dubious and impacted by severe pitfalls such as:
- The method of analogues is too little described; for instance, the analogs metric is not clear. Is it based on precipitation only? If yes, the analog’s distance is too strict.
- It is not clear if analogs are sought in an independent period of the validation period.
- The details of the AR model are not described. Other benchmark stochastic models should be tested.
- By forecast rank, authors mean error, so authors should precise that.
The unique figure presented is not fully discussed. There are results which are not understandable neither discussed such as: the bump in rank around the forecast delay 60-70 minutes for D=1; the reason why the predictability is larger when analogous are sought with D=14 than D=7. Authors present a very speculative unproven reason for that: ‘there is a period doubling bifurcation in the precipitation and that a week periodicity, if it exists, could be unstable’.
Citation: https://doi.org/10.5194/npg-2022-9-RC1 -
AC2: 'Reply on RC1', Yoshito Hirata, 08 Apr 2022
Dear Referee #1,
Thank you very much for providing your insightful comments.
We would like to reply your comments as the attached PDF file.
Especially, please let us know the names of some benchmark stochastic models if we should addn them
because we have alreadly included the AR model, the persistence model and the mean prediction model,
which ,we believe, cover the standard benchmark stochastic models.
Yours sincerely,
Yoshito Hirata
-
RC2: 'Comment on npg-2022-9', Anonymous Referee #2, 23 May 2022
This paper illustrates how conditioning a nowcast precipitation prediction on the calendar day improves a precipitation score.
The poor English syntax and grammar make the manuscript very difficult to follow. The lack of clear and detailed explanations of what is done make the manuscript impossible to assess. For example, the annex is not called in the main text and it does not clarify anything on the procedure that is used by the authors.
From what I see, I do not see how the paper is relevant in NPG, as I do not see a real conceptual innovation (the only innovation appeared in a paper already published by the authors).
Therefore, my appraisal of the paper is based on a guess of what was done to obtain the results.
Major points
The authors use time series with a time increment of one minute. Therefore, not only there is a seasonal cycle, but the time series also contain a diurnal cycle. If there is any cycle in the data, a Fourier transform should be able to detect it. The results reported in Figure 1 do not suggest any type of periodicity.
The methods section does not state how precipitation is predicted (e.g. what model?). Even the AR prediction is not clear. How are the authors certain that they do not over fit the data?
When they use the term “improve” (e.g. in the title), they should state with respect to what? The improvement over operational nowcasting from meteorological institutions should be demonstrated.
I feel that the reported result (better 2h forecast when taking D=14 day prior information) is only valid for the statistical scheme alluded to by the authors. Nothing proves or even suggests that this would hold in a “regular” nowcasting meteorological forecast.
The main result of the paper is based on Figure 1. But this figure does not prove anything, in particular for D=14. The authors have not tried other values of D, in particular larger values. The seasonal dependence is not discussed or even assessed. Why is there a “bump” for D=1? Precipitation differences of 0.006 mm (maximum value of the vertical axis in Figure 1) are not measurable by meteorological instruments. Therefore, the apparent minimum for D=14 cannot be measured in practice. This minimum of mean absolute error might not even be statistically significant (and it is obviously not physically relevant).
The right way to assess forecast schemes is to use cross validation procedures, i.e. at least by considering a training period and a separate validation period. Therefore, it is not even clear that the reported result is actually true.
Specific comments
Abstract: is precipitation dependence a weather variable? (or what is weather variable, and how do the authors define “precipitation dependence”?).
l. 15: Why and how the uni-modal relation (whatever that means) of aerosols and convective energy (why convective energy) is connected to the scattering and absorption of solar radiation?
I think that the authors miss the main point of predicting precipitation, as they treat zero values in the same way as non zero values.
The methods section is inappropriately unclear, especially for a journal like NPG. The first paragraph of section 3 should be in the methods section. The AR model is not defined properly. An order of 120 sounds like overfitting. Precipitation is not Gaussian, especially at minute time scales. An AR model is an obvious bad choice.
Why don’t the authors consider the hour of the day when they condition the forecast? They might avoid an aliasing phenomenon that could explain a fortnight conditioning.
The first paragraph of section 4 is incomprehensible, and is not related to analyses of the paper.
The Appendix section is not really informative on what is done in the forecast.
Conclusion
I cannot recommend the publication of this manuscript in NPG.
Citation: https://doi.org/10.5194/npg-2022-9-RC2 -
AC3: 'Reply on RC2', Yoshito Hirata, 30 May 2022
Dear Referee #2,
Thank you very much for reading our manuscript critically and providing your valuable feedback.
We have examined all of your comments and are replying to you as attached.
Overall, we believe that your valuable feedback helped us to strengthen our findings
because we could show that we can reproduce the similar results with 20-dimensional usual delay coordinates,
a gold standard in the field of nonlinear time series analysis.
Thus, it would be great if you read our response carefully and check that we have overcome all of your concerns.
Again, we deeply appreciate your critical comments, which helped us to demonstrate our findings in the different way.
Yours sincerely,
Yoshito Hirata
-
AC3: 'Reply on RC2', Yoshito Hirata, 30 May 2022
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 1,075 | 166 | 42 | 1,283 | 79 | 28 | 26 |
- HTML: 1,075
- PDF: 166
- XML: 42
- Total: 1,283
- Supplement: 79
- BibTeX: 28
- EndNote: 26
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1