Welcome, Guest
Username: Password: Remember me
General topics
  • Page:
  • 1
  • 2

TOPIC: Visibility in HARMONIE-AROME

Visibility in HARMONIE-AROME 3 years 1 month ago #1737

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
Here at SMHI the aviation forecasters often complain about the visibility forecasts from AROME compared with HIRLAM. I am wondering what your experiences are elsewhere? Do you use the visibility calculations in "gl", or do you do something different?

I wonder since we have more hydrometeors in AROME than in HIRLAM, maybe the same method should not be applied in the two forecasting systems? Or is it the input to the visibility computations that is poor? I saw in the gl computations some assumptions for graupel have been made, has this been updated after we introduced the OCND2 scheme in the reference?

Thank you!

Visibility in HARMONIE-AROME 3 years 1 month ago #1738

  • Sander Tijm
  • Sander Tijm's Avatar
Hi Lisa,

are there specific conditions that the forecasters complain about? Is there too much fog? Is the fog too dense? Are there too many cases with intermediate (1-10 km) visibility? Is there too much low visibility under high wind conditions? These are all things that I have noticed and sometimes they have different causes.

In the gl calculations of visibility not only the impact of hydrometeors (clouds and precipitation) but also aerosols are taken into account. That may also have an impact on the complaints as the aerosol concentration is constant in the model while it is very variable in reality.

So many questions, the answers depend on the conditions that are most problematic.

Visibility in HARMONIE-AROME 3 years 1 month ago #1739

From the summary report on interaction with users in October there was general consensus in several HIRLAM institutes that at too small visibility is forecasted in the HARMONIE, but that is mainly based on CY38. At DMI I think the experience has changed with CY40. It is now often seen that HIRLAM produces more fog and less visibility than HARMONIE. - When there are problems it is almost always coinciding with challenge of predicting cloud base correctly.
I will check more details tomorrow on formulations and discuss with forecasters during our weekly meeting.

Visibility in HARMONIE-AROME 3 years 1 month ago #1740

Hi.

Couple of years ago I tried to understand how the visibility algorithms both in Harmonie (gl) and Hirlam work. As a side product I put together following slides on methods used (could be useful to others as well):

www.dropbox.com/s/8e630at9virv60f/Visibi...mela_update.ppt?dl=0

I think Harmonie visibility method is dominated by the quality of input data, i.e. lowest model lever hydrometeors. As Bent said, the main complaint from our forecasters has been too small visibility (in 38h1.2). Many of the cases are related to overestimation of lowest model level clouds.

About the graupel assumptions... I don't have any knowledge if anyone has tuned the graupel part of the visibility algorithm after introduction of OCND2. At the moment it not the most dominating hydrometeor type in the method (see slide 11).

Visibility in HARMONIE-AROME 3 years 1 month ago #1741

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
Hello, thank you very much for your comments, Sander - I will have to get back to you on your questions, but during our meeting yesterday it came up again for cycle 40, and for the short time it's been operational it seemed that they were even more disappointed in the visibility forecast than in 38h1.2 (still too low visibility), but it can of course be due to the weather situation that we have had. There are no reports yet of excessive fog over sea which we saw during this period quite a lot last year, so the problem is probably mainly over land.

Sami - thank you very much for the presentation, it is very useful. If I understand correctly, the assumption of aerosol is maybe not so important either?

Visibility in HARMONIE-AROME 3 years 1 month ago #1743

Visibility in HARMONIE:

Dear all,

Thanks for all the relevant input, I have some more information or views:

1) Interaction with forecasters: It is fine to have discussions in FORUM, but in addition, in order to utilize information from forecasters, e.g on visibity forecasts, the exchange of information needs to be quite precise/quantitative, e.g. what was the problem identified (visibility –too high/too low) , at what time did it occur and where geographically , preferably also HARMONIE forecast initial time … ). This type of information should hopefully in the future be communicated via the USER contact person in each institute and show up in the quarterly reports that I produce , but this means, in my view, that each institute should try to establish a kind of forecast LOG that can be used to extract information when needed (at least this would be helpful)

2) Visibility routine: At DMI we are using GL-processed visibility: the visibility routine originates from DMI, has update from SMHI (Esbjørn) and from KNMI I believe ( Sander).

3) Example of visibility forecast: This night , e.g 25/11 00 UTC there was quite low visibility over major parts of Denmark. HARMONIE did forecast this reasonably well 12- 30 hours before occurring, but in the first six hours there was a significant over prediction of visibility (too large values of visibility), i.e. “loosing the fog” at short forecast ranges (spinup-problem again). We will continue to follow whether the spinup issue is repeated in other events and from objective verification that we do on visibility. HIRLAM at DMI tends to produce too much fog /too low visibility

Best regards
Bent

Visibility in HARMONIE-AROME 3 years 1 month ago #1744

  • Sander Tijm
  • Sander Tijm's Avatar
Dear all,

here is a small study into the visibility verification for the Autumn of 2016 over the Netherlands and surrounding North Sea.

I have made distributions for the observations plus the HARMONIE version 36h1.4 and 38h1.2+ (the + means that the latest vertical diffusion updates are included in this HARMONIE version, it is therefore quite close to the 40h1.1 version). I have binned the forecasts into 6-hour bins (so all forecasts from +1-+6 are put into the first row) and I have chosen the visibility bins into a logarithmic scale with the boundaries being put at 10**(k/3).

The two tables below show that HA36h1.4 has a significant overestimation of cases with a very low visibility (too dense fog) and an underestimation of the cases with 100-1000 metres. There seems to be a slow decrease in the number of cases with fog as a function of forecast length.

HARM 36h1.4
MIN VIS 22 46 100 215 464 1000 (m)
OBS 0 0 38 282 358 414
6 0 209 88 60 114 232
12 0 177 107 68 71 196
18 0 207 110 61 74 218
24 0 238 128 71 94 212
30 0 200 167 70 90 204
36 0 164 131 98 88 225
42 0 123 118 91 96 178
48 0 86 83 50 82 167

Version 38h1.2 has an overestimation of the number of cases with very low visibility, although that is much less than HA36h1.4. In addition there is a much larger underestimation of the cases with fog than with 36h1.4. What can be seen also is the observation of Bent that close to the analysis time there is less fog than later in the forecast. Later in the forecast there is no decrease in fog cases as seems to be the case in HA36.

HARM 38h1.2+
MIN VIS 22 46 100 215 464 1000 (m)
OBS 0 0 38 282 358 414
6 0 28 28 28 48 141
12 0 27 33 30 34 136
18 0 35 41 26 49 126
24 0 60 54 30 32 130
30 0 68 47 21 38 134
36 0 86 56 38 31 102
42 0 92 51 35 40 162
48 1 108 56 50 54 183

Conclusion: HA38 (HA40) has less fog than observed and the underestimation is larger close to the analysis time.

Visibility in HARMONIE-AROME 3 years 1 month ago #1745

Hi Sander ,

Many thanks for your fog forecast verification study, - I think it agrees well with my impression from CY40 forecasts lately at DMI. I suppose the cloud prediction group needs to continue thinking and discussing how to reduce the systematic under/overforecasting of visibility/fog

Bent

Visibility in HARMONIE-AROME 3 years 1 month ago #1746

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
Thank you Sander! I agree with Bent that it seems to be in line with what we have seen, and the problem is probably not the visibility post-processing as much as the input to the post-processing.

Visibility in HARMONIE-AROME 3 years 1 month ago #1747

  • Sander Tijm
  • Sander Tijm's Avatar
  • OFFLINE
  • Administrator
  • Posts: 24
  • Thank you received: 2
Dear all,

I think that there are three problems that need to be separated.

- when there is fog, the visibility seems to be significantly underestimated. This may be a problem of the postprocessing as we use one single relation to describe the relation between the cloud water concentration and the visibility. In reality this is dependent on the cloud water concentration and the cloud droplet number. When fog forms there are many small droplets. When the fog becomes older, the droplets become larger. When there are fewer, but larger droplets, the visibility should be higher for the same cloud water concentration.

- it looks like HA38/HA40 underestimates the number of fog cases. Maybe the model is too dry or there is too much entrainment of dry air from above the fog layer? Here we have to take into account that the visibility observations are done at a level of 2 metres, whereas the lowest model layer is 20-25 metres thick. So shallow fog banks that may cause the visibility to become lower than 1000 metres are seen as hazy conditions in the model, as only a part of the lowest model layer is saturated. So what the model should look like is quite difficult to determine until we have a lowest model level at 2 metres.

- there seems to be a significant negative impact of the data assimilation on the ability of the model to forecast fog. There is almost a doubling of the fog cases going from +1-+6 to +43-+48 hours forecasts. This may be a result from the synoptic scale impact of radiosonde observations. There is no input to the atmospheric model from the temperature and dewpoint, so the model does not know from SYNOP if there is fog or not. Radiosondes taken far away may have a drying impact on the entire atmosphere close to the surface.

I hope this helps a little in attacking the problem from the right angles.

Visibility in HARMONIE-AROME 3 years 1 month ago #1749

Hi Sander and others,

Thanks for the latest input Sander, I agree with your points which
may serve as som sort of guideline for further Work, -

I found an article with a visibility parameterization depending on both cloud liquid Water and on droplet size, e.g. JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY, vol. 45 1469 - 1480, 2006
These types of parameterizations may offer a framework for improvement
when it is the parameterization and not the input which is the main problem

We also need to be aware of the very shallow fogs over land below the lowest model level. At DMI we will soon take into account all the road weather stations T2M , TD2M (many stations in DK) in the hope that we can analyze relative humidity better close to the surface

Best,
Bent

Visibility in HARMONIE-AROME 3 years 1 month ago #1756

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
Dear all,

I attach some verification of cloud cover that Ulf Andrae has done for a summer, winter and fall period over the MetCoop domain:

Summer cloud cover
bias_rmse_cloud.png

Summer visibility
bias_rmse.png

Winter cloud cover
cloud_cover_winter_rmse.png

Winter visibility
visibility_winter_rmse.png

Fall cloud cover
cloud_autumn_rmse.png

Fall visibility
visibility_rmse.png


Even though there is a reduction of a positive bias in cloud cover from cycle 38h1.2 to cycle 40h1.1, the bias in visibility appears almost unchanged.

Here is an example of the skill scores in visibility and cloud cover for winter:

winter_clouds_skill.png


winter_visibility_skill.png


The skill score of cloud cover is improved, but the skill score of visibility is reduced somewhat.

Now I reckon that cloud cover is not a direct input to the visibility computations, but if we get less over prediction of low level clouds and fog, I would expect it to be reflected in the visibility scores?

Regarding the spin-up, it is a problem that quite many people are currently involved in trying to understand better from many directions.

Visibility in HARMONIE-AROME 3 years 1 month ago #1757

  • Sander Tijm
  • Sander Tijm's Avatar
  • OFFLINE
  • Administrator
  • Posts: 24
  • Thank you received: 2
Hi Lisa,

interesting results, especially the seemingly contradictionary result for the cloud cover and the visibility.

The scores you see here are not saying too much, I fear. The visibility in HARMONIE is dependent on two inputs: the water in fluid or solid form (cloud water, cloud ice, rain, snow, graupel) and the relative humidity.

The concentration of water and ice (different relations for all the parameters) determine the extinction from the hydrometeors, the relative humidity in combination with a fixed number of aerosols (different over land and over sea) gives the "background" visibility.

The average visibility depends most on the cases without cloud water/ice/rain/snow/graupel. So if you want to see what is the problem, an underestimation of the fog cases, or an underestimation of the visibility under normal conditions (aerosol concentration too high, too high relative humidity) you should look for the hit rate, miss rate and false alarm rate for e.g. thresholds of 1000 metres, visibilities between 1000 and 10000 metres and visibilities above 10000 metres.

One other interesting measure may be the frequency bias for different visibility thresholds. I think that that will learn you much more about what is happening than the verification plots you show above.

One other interesting thing to do may be to make a division between the observations made at sea and over land. I think that that may help in objectively finding the weak spots of the model.

Visibility in HARMONIE-AROME 3 years 1 month ago #1758

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
Hi Sander!

Do you know if it is the 2m RH, or is it the whole profile that acts as input. We know that there is a quite huge positive bias in 2m RH, and Patrick Samuelsson has showed that with 2 patches this bias is reduced, in particularly in spring, so I will see if I can do some visibility verification on his experiments.

I'll get back to you on the other plots! Good ideas.

Lisa

Visibility in HARMONIE-AROME 3 years 1 month ago #1759

  • Sander Tijm
  • Sander Tijm's Avatar
  • OFFLINE
  • Administrator
  • Posts: 24
  • Thank you received: 2
This is the calculation of the visibility in visibility.f90 (in gl). The first visibility is calculated based on the cloud water/cloud ice, rain, snow and graupel. The second one is based on the aerosol concentration and the relative humidity at 2m.

In the underestimation of the visibility there are two factors. The first one is the fixed aerosol concentration, which may be on the high side (?), the second one is the relative humidity.

The first one could be improved by importing a forecast of aerosol concentration in the postprocessing (e.g. from GEMS?). The second one may be improved by Patricks work. Concerning the experience of the forecasters there should also be a significant positive bias in rh2m in the seasons where the visibility is underestimated strongly. And if this is the case more in 40h1.1 than 38h1.2, then there should be an increase in the rh2m bias going from 38h1.2 to 40h1.1.

Lines 333-345 of visibility.f90
vis(:,:) = -1000._jprb*log(0.02_jprb)/(ec(:,:)+ei(:,:)+er(:,:)+es(:,:)+eg(:,:)+0.013_jprb)

rh2m(:,:) = MIN(1._jprb,rh2m(:,:))
rh2m(:,:) = MAX(0.001_jprb,rh2m(:,:))

WHERE(ec < eps .AND. ei < eps .AND. er < eps .AND. es < eps .AND. eg < eps)
zccn = zccn*0.2_jprb
cwm = -0.2618e-9_jprb*zccn/(DLOG(rh2m)-0.00001_jprb)
vis = 0.000978_jprb/((LOG(zccn)+0.25_jprb)*cwm**0.6666667_jprb)
vis = vis*1000._jprb
END WHERE

vis(:,:) = MIN(vis(:,:),50000._jprb)

There is one small improvement that could be made in visibility.f90. Now the visibility based on aerosol concentration is calculated only where the microphysical parameters are smaller than eps. But this gives some strange lines in the plots (see the inserted plot)
HA38_20_NL_020.png


These lines could be removed by calculating vis based on the microphysical parameters, a vis 2 based on the aerosols and then taking the minimum value of both (with the maximum of 50000).
  • Page:
  • 1
  • 2
Time to create page: 0.112 seconds