Welcome, Guest
Username: Password: Remember me
Forum on HIRLAM/HARMONIE Observation Handling
  • Page:
  • 1
  • 2

TOPIC: Overall tuning of Harmonie DA

Overall tuning of Harmonie DA 6 years 1 month ago #1140

Gert-Jan Marseille wrote on the 09/26/2013:

I have done a couple of Harmonie experiments, with different
observation settings, and analyzed te results. Main conclusions so far:

1. 3d-var with a 6-hour assimilation window is too large. A 3-h
window is recommended, in particular for extreme weather events with
strong winds.

2. DA does not improve Harmonie forecasts; simply downscaling gives
forecasts of the same quality, on average. This conclusion holds for
assimilating conventional observation (TEMP/AMDAR/SYNOP). Introducing
scatterometer (QuikScat/ASCAT) in addition does not improve the
forecast quality. Overall, ECMWF outperforms Harmonie.

The 2nd conclusion is disappointing, but I discovered some issues
that can explain the so far disapponting results from Harmonie + DA.

To be complete. DA does not detoriate the analyses and forecasts, so
far it does not improve forecasts relative to downscaling only (doing
no DA).

1. The B-matrix is based is based on a 6-hr first-guess (FG) rather
than a 3-hr FG. In fact it is based on the NMC method applied to FC+9
and FC+3.

2. Looking at (o-b)/(o-a) statistics, I find that observations get
too much weight in the analysis. This is explained by too large B
errors (item 1).

3. Relative poor quality observations pass quality control and are
used in the analysis. QC should be more stringent.

4. The observation density of observations used in the analyses seems
too large which may cause overfitting, in particular in combination
with too large B errors.

My plan to improve things is to first generate a B-matrix based on
3-h FG. Next, I want to look at the B-matrix length scales similar as
we did in V(H)AMP. Once I know the latter, I have a feeling on the
density of observations to be used to avoid overfitting.
Alternatively, I could inflate the observation errors.

I think it would be great if we could
collaborate and share expertise and results.
Last Edit: 6 years 1 month ago by Jelena Bojarova. Reason: message was not complete

Overall tuning of Harmonie DA 6 years 1 month ago #1141

Harald Schyberg replied on 09/29/2013

" A 3-h window is recommended, in particular for extreme weather events with strong winds "
Good to hear this, as HARMONIE reference runs in HIRLAM and our
operational setup planned with Metcoop-AROME (the Swedish-Norwegian
implementation going operational next year) will be with 3-h cycling.
Years ago with HIRLAM results were worse with 3-h than 6-h cycling. A
suggested explanation was that surface obs (which dominated at
intermediate termins) added errors at higher levels due to inadequate
structure functions. It might be that observation coverage/observation
usage capability above surface has improved since then.

"DA does not detoriate the analyses and forecasts, so
far it does not improve forecasts relative to downscaling only (doing
no DA)".

The impact of DA and its duration of course depends on domain size.
Information from lateral boundary forcing faster dominates over
information from initial state for small domains. How large was you
domain?

The impression I have from Meteo-France is that their Arome-France
implementation benefits a lot from data assimilation. Could it be due
to larger domain size, or use of more observations, better tuned
system, structure functions or something else?



"Relative poor quality observations pass quality control and are
used in the analysis"
.

I copy this email to Roger and Ole, as they have some hands-on
experience/knowledge with bg error statistics and obs error statistics
used (more than me). I hope they can comment on your findings, what B
statistics they have used, and on whether they see the same thing.


"The observation density of observations used in the analyses seems
too large which may cause overfitting, in particular in combination
with too large B errors" .

Yes, this could be a reasonable hypothesis.

Overall tuning of Harmonie DA 6 years 1 month ago #1142

Gert-Jan Marseille replied on the 30/09/2013

Ad told me that ECMWF had the same problems with inadequate structure
functions back in 1993 when he started assimilating scatterometer
observations.

The size of my domain is 800x800 or 2000x2000 km.

You mention that the Meteo-France Arome-France implementation benefits a
lot from data assimilation. It might be true that their system is better
tuned. That is the reason I started looking at data usage and B-matrix
implementation in our Harmonie implementation. With respect to the
latter I ran a single obs. experiment for temperature at 500 hPa. The
figure (3rd page of attached postscript file) shows that the analysis
increment is non-isotropic with length-scales (HWHM) of about 80 km
(meridional) and about 140 km (latitudinal). That seems pretty large for
a 2.5. km meso-scale model? (The length scales you (and Andras)
calculated for the ECMWF model are only about twice as large, but with
substantial coarser ECMWF grid size.)

The next step is to generate similar length scale plots for a single
u-component observation, but this experiment failed and I hope Roger can
help me here.

Anyway first conclusion here is that assimilating AMDAR at about 10 km
spacing and given B-matrix length scales in the order of 100 km will
result in overfitting when no error inflation is applied. Do you agree?
Attachments:
Last Edit: 6 years 1 month ago by Jelena Bojarova.

Overall tuning of Harmonie DA 6 years 1 month ago #1143

Harald Schyberg replied on the 30/09/2013

- We use "large scale mixing" with the ECMWF model in our AROME implementations. Experience shows that including information from ECMWF like this is important to bring up the quality. This should improve the quality vs ECMWF, but probably does not help when comparing assimilation vs no assimilation. Ole was not sure whether that was default in the HARMONIE version you have been using, so maybe you should check whether that has been used.

- It is possible that our B matrices planned for MetCoop are based on ensemble data and not on the NMC method. Magnus surely knows this. That will give sharper structure functions (smaller length scales).

- There are questions on using NMC method with as short ranges as +3h and + 9h. Analysis increments might not have propagated long enough to make differences between the two runs in unobserved areas.

Overall tuning of Harmonie DA 6 years 1 month ago #1144

Roger Randriamapianina replied on the 4/10/2013

It was nice to read all your comments and suggestions to incorporate
the problems reported by Gert-Jan. It will be difficult to answer
properly to all the raised questions, but I'll try to share my point of
view to the discussed issues.

-- Assimilation with the Cy37 in Harmonie: If you did not do changes
related to the settings related to the upper-air analysis, which is 3DVAR,
then more probably, you have also done your experiment with the
"large scale mixing". This option now is the default in Harmonie.
-- Background error statistics: Indeed NMC method, and the way it
was estimated cannot provide optimal statistics for a 3-h cycling.
The results of your single obs indeed support this idea. As you were
already informed, Malte have done single wind obs assimilation with
the MetCoOp statistics. I think the results were "good" in term of spread
inside the 3D domain. Here, I would indeed appreciate Jelena's suggestion
to compare, again, all the available statistics. I remember Nils showed
very interesting comparison results that time.
-- Correlated observations, thinning issue: Personally, I do not have
experience on using R matrix taking into account observation spatial
correlation errors (i.e. using not only a diagonal matrix). But, if I
remember well Mike Fisher and Gabor Radnoti have done few year ago
similar exercise, and in my memory they did not get significant impact.
We can try to find their report, of course. So, I'm afraid not to get
significant impact of such a tuning without tuning in the methodology
itself. We can have this topic for the Copenhagen meeting for example.
About Jelena's comment on the way the AMDAR data are used now in Harmonie
(25 km thinning), I have made with the ALADIN/HU (11 km hor. res.) a
more aggressive test with higher density. I still have those results if
someone is interested. As far as I remember, the results "were not so
bad". To be honest, I started to think about repeating the same test with
the AROME settings now. I need to check, but I think we have less
thinning distance in the AROME-HU.
-- Need of retuning the system after scientific updates: I completely
agree, this need to be done after "serious" scientific update, but we
always have problem to do coordinated studies (I think it's faire
to recognise this and may be try to find better solution in the future).
METCoOp is very well organised, so I do believe that at least they'll
do similar tuning...
-- We had the idea to have dedicated group for different topics, maybe
we can strengthen this process using distance meetings (for example using
google hangouts, or similar tools, I tried to reach few colleagues
through this tool, and all trials were successful).

As I already mentioned above, let bring these issues to the meeting
in Copenhagen.

Overall tuning of Harmonie DA 6 years 1 month ago #1145

Jelena Bojarova replied on the 30/09/2013

Regarding the structure functions, the easiest way to judge how reasonable structure functions are is to compare them to the structure functions derived for other domains. You may already have a report from Nils with the comparison of structure functions for different HARMONIE AROME domains made 2-3 years ago. We could even repeat the exercise and compare the structure functions for the most actual currently used AROME HARMONIE domains if You send the "background structure files" to us.

The ensemble based methodology to generate structure functions for HARMONIE exist on the hirlam.org, and it was at the latest used by Tuuli Perttula at FMI in CY37. It is a half manual methodology and it is on our to-do list to improve it and make it integrated into the system. Maybe we could ask Roger to help here. I think a number of adaptations will be required to run it with CY38. Jan Barkmeijer has experienced a problem to apply it for so large domain as ECJAN, and he has used software from MF to generate background statistics.

FC_9-FC_3 are certainly not representative for 3h forecast errors. The derived structure functions will depend on what initial field has been used to generate the FC forecasts. If for example the initial field is a pure downscaling of the boundary fields then one cannot expect any reasonable structures on the small scales.

Shu Zhuang at DMI has made a study of the structure functions for HARMONIE domains dependent on various types of climatology (seasonal, duirnal, weather regime; now he is deriving boundary layer stability dependent (Richardson number) structure functions). I think he has a publication on this topic.

It is very difficult to improve the ECMWF forecast using conventional data because the global analysis will always be superior to the LAM analysis. One can only improve the forecast quality by introducing more data more frequently and by improving the structure functions. For low level winds it is in particular important do not propagate increment when it is not needed. Another important topic to consider are the correlated observation errors. One either should account for the correlated observation errors (this is not done currently in HARMONIE) or one should apply so heavy thinning that the correlated errors are excluded. This is what lies behide heavy thining distances used currently for AMDARS for example.

For scatterometer data the correlated observation error are even large problem. QC control procedure is performed in several steps. First ""gross-error" check is done. Then observation is checked on the consistency with the surrounding observations. In case of correlated observation errors such algorithm is very unefficient in discriminating the errorneous observations. Supper-obbing dependent on the resolved scales is needed, and some work in this area has started at SMHI (it is unsure if there will be funds to continue work next year).


The MF data assimilation suits are indeed much better tuned than the HARMONIE ones. It is a complicated task to tune the HARMONIE suits because of different size, different domains, different data coverage. HARMONIE suit for one domain may require different setting for the optimal performance than the HARMONIE suit for another domain. Tuning must be done locally and I do not think that the generally valid recommendation could be proposed. Our main reason for proposing the dedicated action groups was to coordinate this work among the institute, in any case propose the forum where all could share the experience. One of the proposed topics is the tuning of AMDAR data usage incuding the observation error statistics, quality control, bias correction and the thinning distance. The problem is that people proposed to the action group are the same people fully occupied with other DA tasks... It is very good that Your presentation brought again the light on how important the overall tuning of DA sytem is. Ideally, intriducing any new component in DA system, changing model configuration or model resolution, the tuning of the DA system should be performed and MF is doing so.

Overall tuning of Harmonie DA 6 years 1 month ago #1146

Gert-Jan Marseille replied on 8/10/2013



First, I found that large scale mixing is indeed the default setting in cy37. Not by looking at the experiment settings but by generating energy density spectra. I was quite surprised when I saw the result.

ECMWF
First, have a look at 250hpa_ec.png (attached). It shows spectra for the u (left) and v (right) wind components as obtained from ECMWF model fields at model level 53 (around 250 hPa). For reference the curvatures of k^-3 (2-D turbulence) and k^-5/3 (3-D turbulence) are plotting in gray, k denoting wavenumber in m^-1 along the x-axis. Clearly the 3 curves, representing fc+0, fc+3 and fc+6, overlap as we normally see for ECMWF. In fact the spectrum from fc+240 will also overlap: ECMWF spectra are similar for all forecast ranges. Also clear is that ECMWF does not follow k^-5/3 for any spatial scale. For instance, spectra from AMDAR and Mode-S show a k^-3 spectrum for scales larger than 500 km (wavenumbersmaller than 2e-6) and a k^-5/3 spectrum for smaller scales.

Harmonie cold-start
Next, have a look at 250hpa_ha_cold_start.png showing a similar result for a Harmonie experiment without data assimilation (downscaling only). At FC+00 (red curve), the model spectrum is close to ECMWF (not plotted) as expected. Already after 3 hours (and probably already after one hour) Harmonie has added quite some energy on scales below 250 km (wavenumber 4e-6), not resolved by ECMWF, the curve more closely following the atmosphere k^-5/3 spectrum. Same for FC+06.

Harmonie warm-start
Next, have a look at 250hpa_ha_warm_start.png showing a similar result for a Harmonie experiment with data assimilation of conventional observing systems: TEMP/AMDAR/SYNOP. The curvature at FC+00 (red) shows a strange behaviour, following the ECMWF spectrum for scales larger than 250 km, following the Harmonie spectrum for scales below about 25 km and weird mixing in between. Indeed, the large scale mixing of Harmonie with ECMWF becomes clearly evident here, with a large spectral gap in the Harmonie field for spatial scales between 25 and 250 km.

Spectra from ICMSHANAL+0000 and MXMIN1999+0000 also show the same weird spectral behaviour.

From what I understand the reason to apply large-scale mixing with ECMWF is the superiority of the latter on the large and medium range scales. from this I conclude that you have low confidence on the added energy by Harmonie on the 25-250 km scales. In other words, apparently the added structure does not verify (with observations). But then I would also have no confidence on the smaller than 25 km scales, so why analysing these in DA?

Second, how can we expect the structure functions to be balanced (already mentioned by Ole in hirlam.org/trac/attachment/wiki/Harmonie...entation/lsmixbc.ppt) and correct at all if they were generated without taking into account large-scale mixing?

Third, can we expect substantial positive impact from analysing spatial structures on scales that do not verify, i.e., noise?

After some discussions with colleagues at KNMI I would propose the following for Harmonie 3D-var:
1. Analyse only those scales that verify with observations. This can be tested from (o-b) statistics for a range of truncation wavenumbers applied to b. For small truncation wavenumbers (o-b) will decrease with increasing wavenumber. At a given threshold wavenumber, k_thresh, (o-b) will start to increase, meaning that the additional spatial structures added to b do not verify, i.e., noise.
2. Rather than applying large-scale mixing, truncate the first-guess at k_thresh and analyse the large scales only. I would hope that Harmonie large scales well match with ECMWF! In that case we could use the ECMWF ensemble to generate flow-dependent B structures for use in Harmonie.

With the steps above, we are at least sure to have a better founded DA implementation, including well-balanced flow-dependent B structures.

With the growing number of observations (both spatially and temporal) we can hope to better resolve the small Harmonie scales in the future.

Observation error correlation, thinning, etc. is all very important but starts after having well understood which scales to analyse in DA, rather than trying artificial mixtures.
Attachments:

Overall tuning of Harmonie DA 6 years 1 month ago #1147

Ad Stoffelen replied on the 8/10/2013

"Third, can we expect substantial positive impact from analysing spatial structures on scales that do not verify, i.e.,

In this context, I provided talks a few years ago at NCEP, noise?"ECMWF, UKMO and JCSDA-HFIP Workshop on "Satellite Data Assimilation for Hurricane Forecasting" and at KNMI. Slides 19-27 in the latter contain some thoughts and ideas which you may want to reflect upon, see at www.knmi.nl/publications/fulltexts/kik_lunch.pdf . It addresses the problem of resolved but uninitialized model scales.

I also particularly recommend slide 9 for further reflection.

Overall tuning of Harmonie DA 6 years 1 month ago #1148

Malte Muller replied on 9/10/2013 (to Gert-Jan mail 8/10/2013)


I followed you idea in analysing the impact of 3DVAR on the
wavenumber spectrum of wind velocities. I did a similar analysis
for the METCOOP25B region. The structure functions I am using are
available on ecgate, but actually I do not know how they are generated.
I am using cy37v1.2 and the default setup.

I show 2D wavenumber spectra but I am not detrending nor windowing,
so I think I am loosing information on the very large scales as
you will see in the sharp bend of the spectra at the low wavenumber end.

1. The coldstart (Spectra_radar_blending.tif) downscaled from ECMWF
shows a similar behaviour as you already observed. There is a lack
of scales below the resolution of the ECMWF model.

2. The warmstart (3DVAR and Background is 6 hour forecast) looks
different than yours. I do not see the same loss in energy for
scales between 25-200km. I assimilate conventional plus radar radial
wind data, and the curves look similar if I only use conventional
data.

3. I compare the Background spectra and the analysis spectra (of u wind) in the plot
Spectra_radar_BGvsAN_3DVar.tif directly. The differences are fairly small,
and I believe this is something we would expect?!

So it seems that either the reason for the discrepancy of our findings is
in the structure functions or that we do not have the same switch for spectral
mixing (on/off)?
Last Edit: 6 years 1 month ago by Jelena Bojarova.

Overall tuning of Harmonie DA 6 years 1 month ago #1149

Jelena Bojarova replied on 10/10/2013

I attach here the similar plot done by Nils with the analysis of the spectra in HIRLAM and HARMONIE ( this is a part of work done in collaboration with Ljubljana University).

Regarding the structure functions for MetCoOP domain, they are generated from the ensemble of 3 hour forecasts (downscaling of boundary perturbations) sampled over different seasons of the year. Such sampling was done to assure that the statistics would be as representative as possible. So I think it is indeed different methods were used to generated the structure functions for ECJAN and MetCoOP domain.

For any LAM domain the large scale error should be taken into account in some way and their impact is mostly on long and medium range forecasts. Large scale mix is an engineering solution which seems to work quite well. However, I do not think it is reasonable to use large scale mix option when one calculates structure functions. Large scale mix has a quite drastic impact on spectra.
Large scale error constraint in the cost function term is another way to take into account the impact of large scale error. The implementation of the Jk term is very different in HIRLAM and HARMONIE systems, and it seems that the approach chosen in HIRLAM is much more powerful in comparison to one chosen in HARMONIE. The main difference in the implementation is that in HIRLAM one takes the Jk term as an extra term in the formulation of the background constraint modifying the preconditioning at the same time and is imposed on vorticity only, when in HARMONIE the Jk term comes as extra observation of the whole host model state. We should indeed discuss if one need to refine the implementation of the Jk term in HARMONIE. Large Scale Mix is implemented by Ole Vignes.
I am not sure how it is exactly implemented, but I think changing the size of domain one should modify some parameters of the scheme. Different size domain will resolve different spectra...



Regarding the question of the structure functions and the reserved scales in HARMONIE , we need indeed invest more time and efforts in addressing this area. Some work has been done at SMHI by Tomas Landelius mainly. He has been looking on the appropriate scale for super-obbing dependent on the what scales are resolved by the model. This work went mainly on external funds and it is not clear if there will be funds to continue this work next year. We in HMG are finilizing the update of the rolling HIRLAM-B/ALADIN plan with tasks foreseen in 2014. Please tell me if You consider You will have possibility to look into this important topic next year.

The formulation of structure functions and the resolved scales in the analysis are mutually impacting issues. Only by using high-frequency high-resolution data and a more elaborated structure function we can add value to the analysis provided by the global model. I do not think one can expect improvement of the analysis using the conventional observations.
Attachments:

Overall tuning of Harmonie DA 6 years 1 month ago #1150

Gert-Jan replied on the 13/10/2013 (to jelena 10/10/2013)

I attach here the similar plot done by Nils with the analysis of the spectra in HIRLAM and HARMONIE

The Harmonie spectral gap is not well visible here. Which Harmonie model version has been used? I am wondering why the 0h spectrum differs from the 3h on the small scales? Malte and my spectra show that the 0h/3h/6h spectra are similar on scales smaller than 25 km.
What about the amplitudes of the spectra. I can not believe that Hirlam and Harmonie 6h have similar energy on scale around 10 km. I compared Hirlam and ECMWF spectra and they are very similar. Both ECMWF and Hirlam show negligible energy on these scales.


However, I do not think it is reasonable to use large scale mix option when one calculates structure functions. Large scale mix has a quite drastic impact on spectra.

So we accept that model spectra and structure functions are not in agreement in DA?

Regarding the question of the structure functions and the reserved scales in HARMONIE , we need indeed invest more time and efforts in addressing this area. Some work has been done at SMHI by Tomas Landelius mainly. He has been looking on the appropriate scale for super-obbing dependent on the what scales are resolved by the model.

Very useful indeed. At KNMI we do a similar effort with Mode-S data.
My work is funded by the EU and my contribution to the project finishes around mid 2014.

Only by using high-frequency high-resolution data and a more elaborated structure function we can add value to the analysis provided by the global model. I do not think one can expect improvement of the analysis using the conventional observations.

I think this is too pessimistic. De Haan and Marseille et al. (2013) demonstrated positive impact of ASCAT in Hirlam. De Haan and Stoffelen (2012) demonstrated positive impact of Mode-S in Hirlam RUC (1-hour cycle). As said above, The posititve impact was mainly on the large scales, since Hirlam does not resolve scales below 200 km. The reason that Mode-S showed positive impact was because of its thinning to 50 km spacing (rather than using the raw 1-km spacing). So, I think Harmonie large scales can also profit from conventional aircraft obs., if they are used correctly.

I agree that resolving the small scales really requires an observing network on those scales and more elaborated structures functions. The former seems an utopy. Radar, SCAT and Mode-S data will never be sufficient to cover the complete 3-dimensional Harmonie domain. But we can still use these to improve the Harmonie scales of around 100 km and add to ECMWF here, because of our 3-h cycling rather than the 12-h of ECMWF.

Overall tuning of Harmonie DA 6 years 1 month ago #1151

Ad Stofellen replied on the 15/10/2013

Some observations:

1) HARMONIE spectra are fundamentally different from ECMWF and HIRLAM spectra; representation and scale arguments need to be reconsidered;

2) The forecasted ECMWF large-scale flow is a priori known in experiments and operation; why not nudge the large scales to ECMWF at every time step to prevent spin-up effects at the analysis cycles?

3) The linear filter has a very limited range between 50 en 70 km; KNMI analyses show that the ECMWF spectrum is k-3 while HARMONIE is k-5/3 from 500 km down to 25 km; the latter verifies (above 25 km) with collocated scatterometer, buoy and aircraft spectra; so, we estimate the effective resolution of ECMWF to be at least 10 times grid distance; so, no reason to trust ECMWF scales below 200 km!

4) How was the linear filter range derived? Based on KNMI spectral analyses and collocated comparison to available observations, I’d suggest mixing on 100-1000 km scales (note 1) though)

5) The spectral analyses for Jk appear insufficiently sampled; figure 3 shows wave number 5 vorticity variances of ECMWF about a factor 5 larger than those of HARMONIE; this is in my view totally unrealistic and caused by aliasing; it makes the matrix V rather inaccurate obviously; HARMONIE uses spectral analyses over 750-960 points, but for ECMWF only about 150 are used, which is very limited indeed;

6) I’m not sure how the spectral computations (sampling, windowing) are done, but all steps should at least be identical for all input fields to allow representative comparison and computation;

7) Small-scale structures, edge effects and wave-like disturbances in the HARMONIE model do not appear to verify with observations; this may cause detrimental contributions to o-b which are hard to correct in the analysis step; following the above scale analysis I would try and attempt to correct the 100+-km scales for now in the HAMONIE inner area; this may be achieved in a balanced way I believe;

I’d appreciate further discussion on these issues before experimentation continues, since in the current configuration of Gert-Jan we may not learn too much from his experimentation. Feel free to extend to other HIRLAM scientists.



See also:

Jeuken, A. B. M., P. C. Siegmund, L. C. Heijboer, J. Feichter, and L. Bengtsson (1996), On the potential of assimilating meteorological analyses in a global climate model for the purpose of model validation, J. Geophys. Res., 101(D12), 16939–16950, doi:10.1029/96JD01218.



Marseille, G.J. and A. Stoffelen, 2012: Aeolus Observation Representativeness Error, ESA. Technical Note of VHAMP poject, AE-TN-KNMI-VHAMP-007_v1.0, available from Gert-Jan Marseille.



Jur Vogelzang, How to calculate wind spectra, Satellite Application Facility for Numerical Weather Prediction, NWP SAF, Document NWPSAF-KN-TR-008, Version 1.1, 04-04-2013, KNMI, De Bilt, The Netherlands, research.metoffice.gov.uk/research/inter...ulatewindspectra.pdf

Overall tuning of Harmonie DA 6 years 1 month ago #1152

Jelena Bojarova replied on 16/10/2013

I hope indeed that this discussion will trigger more people to come with the input on these important questions and thank again to Gert-Jan for initiating it.

I do not think one can expect improvement of the analysis using the conventional observations

Yes, my statement was egagerated and not correct. I should add "with 6h hours update frequency"


So we accept that model spectra and structure functions are not in agreement in DA?

Yes, certainly. I think if one is analysing the performance of the DA algorithm one should switch off large scale mixing. It is very different to analyse the impact of the observations if they are overwritten by the structures of the large scale model. Regarding the HARMONIE Jk implementation, it is very different from one chosen in HIRLAM. In HIRLAM Jk term comes as an additional background constraint on the vorticity only and impact even the preconditioning and takes into account the large scale balances. In HARMONIE Jk term comes as an additional observation of the whole host model state ignoring the large scale balances (al least in the implementation what Per Dahlgren has investigated).

I think important outcomes from this discussion is that we have identified several problems which are to be investigated designing the DA system on convection permitting scales

Question one: How one should generate structure functions for limited area model? Host model resolve large scale and is not controlled by the Data Assimilation in LAM. At the same time large scale processes control processes initiated on smaller scale to large extend.

Question two: On what scale it is meaningful to assimilate information. What scales are observable? What scales are described by the model? It is very important message that the it is unrealistic to expect observability up to the finest scales. What scales are described by the model depend on the parameterisation of the physical processes as well. The question of the meaningful scales is a very essential question now for us designing the Hybrid Variational Ensemble data assimilation.

Question three: How do we coordinate in a better way the experimentations carried at different institutes



Two proposal came so far from discussion:
1. Make an overview of the structure functions used at different institutes. There are two methods usually used : the ensemble generation technique and the NMC methods. Both should be applied with care in particular for short forecast lead times. We have too little insight what structures these two methods provides for short lead forecast length and how much structure functions are infected by spin-up problems in particular when downscaling is used.


2. Gert-Jan has proposed the following experiments:
1. Analyse only those scales that verify with observations. This can be tested from (o-b) statistics for a range of truncation wavenumbers applied to b. For small truncation wavenumbers (o-b) will decrease with increasing wavenumber. At a given threshold wavenumber, k_thresh, (o-b) will start to increase, meaning that the additional spatial structures added to b do not verify, i.e., noise.
2. Rather than applying large-scale mixing, truncate the first-guess at k_thresh and analyse the large scales only. I would hope that Harmonie large scales well match with ECMWF! In that case we could use the ECMWF ensemble to generate flow-dependent B structures for use in Harmonie.


I think these experiments are worth of attention. I want only to stress that the verification of model fields against observation is not a trivial procedure when small-scale processes with inherently stochastic structures.
The location of the phenomena can be wrongly predicted even if the phenomena is well described. Another problem that it is not enough to sample an ensemble of several model states in order to obtain a flow-dependent data assimilation. One cannot sample ensemble over too long period: this will destroy flow-dependency. With small ensemble size one cannot describe the forecast error covariance of so huge model state.

COMMENTS ARE VERY WELCOME !!!!

Overall tuning of Harmonie DA 6 years 1 month ago #1156

Question:

Relaxation/nudging towards the host model information high up in the stratosphere is another way to account for large scale error information to some extend. I think this scheme is implemented in HARMONIE (both for AROME and ALARO ?). What are namelist variables controlling it? Does it exist some paper/note describing the scheme?

Overall tuning of Harmonie DA 6 years 1 month ago #1157

  • Ole Vignes
  • Ole Vignes's Avatar
  • OFFLINE
  • Administrator
  • Posts: 36
  • Thank you received: 10
In Harmonie, spectral nudging (coupling) during the forecast is implemented for AROME (not ALARO). In nam/harmonie_namelists.pm you can find, under %arome :

NEMELBC0A=>{
'LESPCPL' => '.TRUE.',
'NEFRSPCPL' => '1,',
'NEK0' => 'ZNEK0,',
'NEK1' => 'ZNEK1,',
'NEN1' => 'ZNEN1,',
'NEN2' => 'ZNEN2,',
'SPNUDDIV' => '0.01,',
'SPNUDQ' => '0.0,',
'SPNUDT' => '0.01,',
'SPNUDVOR' => '0.01,',
},

NEK and NEN are wave number and level limits (low/high).

Regarding LSMIXBC, it is indeed an engineering solution, and was implemented as a test to see whether we could have a cheaper alternative to the re-forecast procedure that is used in HIRLAM. There is no scientific documentation of it except a ppt that was mentioned in an earlier post: hirlam.org/trac/attachment/wiki/Harmonie...entation/lsmixbc.ppt
It was never my intention to include this as part of the common code, but in any case it was included in a phasing at some stage and accepted by Meteo-France, so here we go.

Per Dahlgren compared LSMIXBC and Jk in a MetCoOp study:
metcoop.org/memo/2013/02-2013-METCOOP-MEMO.PDF

LSMIXBC has two tunable parameters, I quote from an earlier mail:
"there are 2 parameters in namelist NEMCT0 that control how LSMIXBC behaves:

ERESBC - resolution (in degrees) of host model, default 0.15.
This controls the weighting of horizontal scales ("cutoff" wavelength),
except that there is no sharp cutoff, but a gradual transition around this wavelength.

VEXPLSMIX - exponent controlling vertical (profile) weights, default 2.0,
which means that the vertical weighting (for a given horizontal scale)
looks like a parabola, weight of host model is 1 at the top and decreases
to zero at the ground. The weights of host model for wavenumber zero at each level
are printed in the output file (NODE_01.001, for Canari) if LSMIXBC=yes."

I think it was a bad choice to set the default ERESBC to 0.15. I totally agree with Ad and others that ECMWF does not resolve scales down to 0.15 degrees well, even though this is their theoretical resolution. Something like 1.5 would indeed make more sense.
  • Page:
  • 1
  • 2
Time to create page: 0.104 seconds