Improving PM<sub>2. 5</sub> forecast over China by the joint  adjustment of initial conditions and source emissions  with an ensemble Kalman filter

Peng, Zhen; Liu, Zhiquan; Chen, Dan; Ban, Junmei

doi:https://doi.org/10.5194/acp-17-4837-2017

Articles | Volume 17, issue 7

https://doi.org/10.5194/acp-17-4837-2017

© Author(s) 2017. This work is distributed under
the Creative Commons Attribution 3.0 License.

https://doi.org/10.5194/acp-17-4837-2017

© Author(s) 2017. This work is distributed under
the Creative Commons Attribution 3.0 License.

Articles | Volume 17, issue 7

Research article

|

13 Apr 2017

Research article |

| 13 Apr 2017

Improving PM_2. 5 forecast over China by the joint adjustment of initial conditions and source emissions with an ensemble Kalman filter

Zhen Peng, Zhiquan Liu, Dan Chen, and Junmei Ban

Download

Final revised paper (published on 13 Apr 2017)
Preprint (discussion started on 26 Aug 2016)

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

RC1: 'Review of “Improving PM2.5 forecast over China by the joint adjustment of initial conditions and source emissions with an ensemble Kalman filter” by Peng et al.', Anonymous Referee #1, 26 Sep 2016
- AC1: 'Response to Reviewer #1’s comments:', Zhen Peng, 02 Dec 2016
RC2: 'review report', Anonymous Referee #2, 19 Oct 2016
- AC2: 'Response to Reviewer #2’s comments', Zhen Peng, 02 Dec 2016

Peer-review completion

AR: Author's response | RR: Referee report | ED: Editor decision

AR by Zhen Peng on behalf of the Authors (02 Dec 2016) Author's response Manuscript

ED: Referee Nomination & Report Request started (05 Dec 2016) by Toshihiko Takemura

RR by Anonymous Referee #1 (17 Dec 2016)

RR by Anonymous Referee #3 (23 Jan 2017)

Suggestions for revision or reasons for rejection

The authors present the results of a forecasting system that assimilates both initial hourly aerosol concentration and emission fluxes in order to improve the forecasting of particulate matter concentrations over China. To evaluate the performance of this system the forecasted concentrations are contrasted on one hand with independent observations not assimilated by the system and on the other hand against a control run without any assimilation and a forecast experiment only assimilating initial conditions but no emissions. The forecast is conduct for all China but a more in depth analysis is conducted in three regions experiencing stronger pollution levels. These three regions are the Beijing-Tianjin-Hebei region, the Yangtze River delta and the Pearl River Delta. The authors present results illustrating that the forecast assimilating initial conditions and emissions performs much better than the control simulation. Performance analysis in the three above-mentioned regions suggests that the system achieves improvements for almost all 48-h forecast in two of them while in the third one the improvement is more limited. Similarly the performance of the joint assimilation compared to the one only assimilating initial conditions shows improvement in two of the regions.

The results presented in the manuscript are interesting, however the authors conduct only a shallow analysis of the results and do not discuss how some of the assumptions made in the system affect the result. Although I recommend this paper for publication I would suggest the authors extend the discussion of the results addressing some of the topics highlighted below. When presenting a new inversion system, in addition of presenting the main results (if it works or not), the limitations of the system and their impact should also be presented.

General comments

The authors assume prior emissions constant in time but it is well known that emissions are not constant throughout the day. Why were emissions considered constant throughout the day and also throughout the week? How much of the improvement in performance of the system comes from this assumption? How much better does the control perform when variable emissions within the day are allowed? Furthermore, the implications of not perturbing emissions of elemental carbon and organic carbon should be included in the manuscript. How does this affect the forecast? How realistic is the result provided by the system with this constrain?

The authors examine first the performance of the system by comparing the analysis of both assimilation experiments (expC and expJ) to the observations and then the forecast. It is interesting to note that when the analysis of both experiments are examined a better performance is obtained in PRD and JJJ when only initial conditions (IC) are assimilated (i.e. expC). However, when comparing the forecasts between both experiments, expJ performs better than the forecast of expC. What are the implications of this result? Furthermore, the authors provide a too simplistic analysis of the performance of the forecast in the three regions. Yes it is true that expJ improves with respect to the control and expC in YRD and PRD, but this is mostly for daytime, during night-time the improvement is very similar in three regions. In YRD, the performance is actually deteriorated during nighttime and in JJJ there is either deterioration or no improvement after 24 hr forecast for both assimilation experiments. Although the authors suggest that this is mainly to a good performance of the model during nighttime, this is not enough I believe. Why is the performance of the control run better during night? Why does the assimilation have so little impact during night? Why should the model have better performance for nocturnal conditions? Was it tuned under such conditions? Do the a priori emissions provided, the ones considered constant, correspond to night emissions? I would suggest the authors spend a bit more trying to address this issue as they have done so far.

If the difference between the control run and expC can be seen as the contribution of assimilating concentrations, can the difference between expC and expJ as the impact of assimilating emissions? If so, is it really worth if to assimilate both? Why wasn’t there and expE conducted where only emissions were assimilated? Figure 8 suggests that in most of the days in the three cities, the fact of assimilating only IC has little impact on the forecast. Figure 9 also illustrates that most of the improvement comes when emissions are assimilated. What if only emissions were to be assimilated, could that be enough? I suggest the authors include a discussion section where this is addressed.

The assimilations system needs further description. The authors describe how the observation error covariance matrix (R) is defined but do not do the same for the background error covariance matrix (Pb). How is Pb defined? The authors should explain this in the manuscript. Furthermore, observations from 77 stations were assimilated and observations from another set of 77 stations were used for verification purposes. However, in the three regions of interest in the manuscript; namely JJJ, YRD, and PRD, it is not clear how many stations were assimilated and how many were used in the verification. This number is provided in the caption of Figure 1 but should also be included in the text. Please also clarify if all these verification stations are used to compute the statistics presented in Figure 9.

Specific comments

Lines 30-31: Acronyms should be defined.

Lines 79-81: Structure of the paper described is not consistent with actual structure of the paper. There are 6 sections in the manuscript and only 5 according to text in last sentence of section 1.

Line 131: Sub index i should be defined. It is clear from the text what it stands for but should be introduced anyway.

Line 132: Why is it t-2 for the emissions and t-1 for the concentrations (line 131)? Is it a mistake and it should be t-1 for both? If not, please explain.

Line 147: Please explain which criteria was applied to define the limits (0.1 and 1.25) to the spread of (Ki,t)inf. How were they defined?

Line 150: Why are the negative values set to 0.001 and not simply 0? Please explain.

Line 322: Remove “which is a limitation of this manuscript”. It is already stated in lines 300 and 301.

Lines 352-356: Explain the criteria used to select the stations that would be used for verification and those used in the assimilation? How many of each are in the different regions. The total number of stations in each region is provided but it is not said how many of them are for validation or verification purposes.

Line 371: Why are hourly concentrations above 800 μg m-3 considered unrealistic? Hasn’t had China intense pollution events where this limit was exceeded in terms of hourly concentration? In any case, this should be argued much better if observations are removed. Also, why are observations where the departure of the ensemble mean of the first guess exceeds 100 μg m-3 removed? What

Line 408: What is the impact of considering that no correlations exist between emission variables. What is the impact on the assimilation and the forecast?.

Lines 460-461: What is it, are the emissions perturbed or not in expC? According to this line not, but according to the statement in lines 450-452, the emissions are perturbed by adding random noise.

Lines 566-570: Where are the numbers in this paragraph coming from? Please explain and present them.

Line 609: Replace “analysing” with “analysis”.

Line 649: What exactly is “dramatic”? How large is that? Please replace.

Lines 1097-1101: Authors should specify if the analysis presented in the figures include all verification stations in each region or only some of them. In addition, authors should also clarify to which dates the analysis presented in the figures corresponds.

Hide

ED: Reconsider after minor revisions (Editor review) (25 Jan 2017) by Toshihiko Takemura

AR by Zhen Peng on behalf of the Authors (07 Feb 2017) Author's response Manuscript

ED: Reconsider after minor revisions (Editor review) (24 Feb 2017) by Toshihiko Takemura

The authors are asked to revise the manuscript carefully according to reviewer's comments (Referee #3) as shown below.

-----

I gave went through the manuscript and the authors made a decent effort to try to address the issues I raised in my comments. I still have one comment though and that is that the authors should be more honest with their results. I’m not saying they’re hiding information or twisting it around but they should acknowledge that for some cases the assimilation just didn’t improve things, I’m talking about the performance of the assimilation in JJJ. There is some improvement in the first 24 hrs but then there is no difference between the control run and the forecasts of both experiments. Considering they’re focusing on the 48hr forecast this is not a minor issue. The authors do mention this in section 5.4 but then in the discussions and conclusions they state “Large improvements were achieved for almost all the 48-h forecasts, particularly in the YRD and PRD. However, relatively smaller improvements were achieved in the first 24-h forecast in the JJJregion, which …”. They focus on the 48hrs in the analysis of the results and then switch to the 24 hr forecast in the conclusions in the conclusions because it did have some improvement. I can understand that the authors want to highlight the things that worked in their assimilation, specially after all the work they have done, which is really good. All I’m saying is that the authors should be honest with the results and if something doesn’t work just say so, it doesn’t make the work less valuable. Based on the results the authors provide, the assimilating simply doesn’t work for the 48 hr forecast in some regions.
What I also find interesting and the authors didn’t address is that based on the analysis runs (table 1) there is not much difference in assimilating IC or IC+emissions (the authors themselves claim the statistics are the same, I would say they are even better for expC in some cases). Anyway, considering that and some of the results in figure 10 (in JJJ and YRD during the night), the improvement of assimilating emissions in addition to IC seems to be not so big or even negligible. I would have expected the authors say something about that as well. Is there really a benefit of assimilating both IC and emissions everywhere? my answer would be no, but the authors suggest that it is the case, at least imply it.
Last thing, one argument used to explain the performance of the system there was the sparsity of the network, but in PRD the network is even sparser and the system performs well so I don’t see the argument. The authors should better explain why in JJJ it is a problem but not in PRD.

Some specific comments (based on the version with the modifications highlighted):
Line 705: it says P25, i suppose it should be PM25
Line 748-749: What do they mean with “Since we did not know the exact station type,….”, Do they mean urban or rural? please reformulate or clarify. In that same sentence the “We” after the , should be “we”.
Lines 1061-1066: As far as I understand the reasons they are providing go in the different direction as what they try to justify. They are addressing the similar performance of the control run and both forecast experiments in YRD during the night. The claim that the prior emissions are larger than the optimized emission during the night and therefore the control run performs worse during night and better during day. What I see in figure 10 is the opposite, the control run performs better during the night than during day. The RMS and bias is smaller during the night than during the day. I would strongly suggest the authors to revise this part.

Lines 1099-1102: I would ask the authors to reformulate these lines, I’m sure it can be improved.

Line 1104: replace “RRD” with "PRD"

Hide

AR by Zhen Peng on behalf of the Authors (28 Feb 2017) Author's response Manuscript

ED: Publish as is (22 Mar 2017) by Toshihiko Takemura

AR by Zhen Peng on behalf of the Authors (22 Mar 2017) Manuscript

Short summary

In order to improve the forecasting of atmospheric aerosols over China, the ensemble square root filter algorithm was extended to simultaneously optimize the chemical initial conditions and primary and precursor emissions. This system was applied to assimilate hourly surface PM_2.5 measurements. The forecasts with the optimized initial conditions and emissions typically outperformed those from the control experiment without data assimilation.

Improving PM2. 5 forecast over China by the joint adjustment of initial conditions and source emissions with an ensemble Kalman filter

Download

Interactive discussion

Peer-review completion

Improving PM_2. 5 forecast over China by the joint adjustment of initial conditions and source emissions with an ensemble Kalman filter