Messages posted by: Stephen Jascourt

SREF (Short-Range Ensemble Forecast system) is changing on Oct 27 - see the COMET Operational Models Matrix and Ensemble Matrix for the updates.

However, NCEP was sending the new data accidentally before this date, as per this October 23 notice from NCEP:

"From 28 September through 22 October, NCO was inadvertently disseminating short-range ensemble forecast (SREF) mean, spread, and probability products to AWIPS from both the production and parallel runs. The two SREF runs were sent out at almost the same time. So for any given cycle, users may have been receiving products from either run, or even a combination of the two runs. The SREF products are now being sent exclusively from production to AWIPS."

Since the SREF changes have demonstrated forecast improvement, hopefully this oversight only helped to improve the forecasts of WFOs utilizing the SREF, although it may have made for some confusing inconsistencies if the old and new were mixed together in the output received.

NCEP lists the following as the changes, but note that new output fields and times are not distributed over NOAAport, instead the existing suite of output parameters and times will be distributed from the revised ensemble:
1) Upgrade WRF_NMM version and increase horizontal resolution from 40km to 32km
2) Upgrade WRF_ARW (aka WRF-EM) version and increase horizontal resolution from 45km to 35km
3) Upgrade RSM version and increase horizontal resolution from 45km to 32km
4) Replace a pair of Eta_sat members with WRF_nmm members
5) Replace a pair of Eta_det members with WRF_arw members
6) Replace a pair of RSM_ZhaoCloud with RSM_FerrierCloud
7) Upgrade BUFR code for WRF_ARW members
8) Adjust IC perturbations by using global Ensemble Transform perturbations for NMM and ARW members
9) Breakout the big BUFR output (with all 1376 stations all together) into individual station time-series
10) Increase output frequency from 3hrly to hourly for the 1st 39hr on grid212
11) Add aviation fields into ensemble products
12) Add wind variance into ensemble products
13) Add Richardson-Number based PBL height to the output products.
14) Add simulated radar Composite Reflectivity and echo top to the SREF output products.
15) Adjust all related aspects of the SREF system such as ensemble product generator and bias correction

Stephen

QuikSCAT winds over water are not being assimilated in RTMA over all domains, as of 20 UTC October 16, due to poor observation quality.

Stephen

The December 2008 NAM changes combined with the March 2008 NAM changes have enabled the NAM WRF-NMM to overcome many of the shortfalls during the initial implementation of WRF-NMM in the NAM slot, replacing the Eta model, back in 2006, finally achieving the success that had been anticipated. The combination of these two excellent implementations places the NAM on an even par with the GFS in synoptic forecast skill all the way through 84 hours.

The March 2008 NAM changes, as well as the December 2008 NAM changes, each separately led to consistently large, positive forecast impacts, much more so than typical of model changes over the past 10 years. While many factors probably contributed to the March model improvement, the largest contribution seems to have come from the gravity wave drag and mountain blocking, which had not previously been included in the Eta model nor its successor in the NAM, the WRF-NMM model. It is explained on the page
http://www.meted.ucar.edu/nwp/pcu2/namturb2b.htm
or you can find it navigating the COMET Operational Models Matrix at
http://www.meted.ucar.edu/nwp/pcu2
scroll down to the Turbulence row at the bottom of the Model Physics section and click on the link in the NAM WRF-NMM column. Then, click on the Scheme Implementation tab at the top of the page and then click on the Topography tab among the expanded choices at the top of the page.

Stephen

A new NAM product, downscaled grids, officially began to be disseminated across NOAAPort in December.
In AWIPS, these are labeled NAMdng, where DNG stands for Downscaled NWP [Numerical Weather Prediction] Grids.
These products downscale the 12-km NAM forecast to a finer grid using the same terrain fields used for the RTMA and using the full set of model fields available at NCEP.

The NAM DNG products appear in AWIPS in both D2D and GFE and can be used to substitute for the "NAM12" GFE products which use downscaling algorithms limited to the vertical levels and parameters of data available in AWIPS.

Details and comparison between the GFE "NAM12" and DNG products are presented on the DNG pages of the COMET Operational Models Matrix at
http://www.meted.ucar.edu/nwp/pcu2/
scroll down around 80% toward the bottom of the page, to the row labeled Postprocessing/Products, and click on the NAM DNG link in the NAM column.
Today, I am still working on the discussion for temperature, dewpoint, and wind - these are coming shortly. However, all background and the discussions for PoP and sky cover are complete, and side-by-side comparisons for temperature, dewpoint, and wind are there already too. I highly recommend making sure that the Java installed in your web browser has sufficient memory to display the comparison loops. Instructions are provided on the web page containing the loops.

Stephen

A major NAM implementation was made in December. Highlights are summarized on the COMET Operational Models Matrix page at
http://www.meted.ucar.edu/nwp/pcu2/
in the "What's New!" section at the top of the page.

The most important change with a consistent positive forecast impact in the 2-3 day time range is "partial cycling". The NAM uses a data assimilation cycle that starts 12 hours before the forecast initial time, such as 00 UTC the previous night for the 12 UTC morning forecast cycle. At the 12-hours-prior time (for example, 00 UTC), an analysis is performed using a first guess plus observations valid around that time. It is this 12-hour-old first guess which has changed. Then a 3-hour "forecast" is run, serving as the first guess to an analysis using more observations at 9-hours-prior (for example, 3 UTC), and repeated in 3-hourly increments up to the time of the forecast cycle (for example, 12 UTC). Thus, in total, there are four 3-hour "forecasts", each serving as as the first guess to the next analysis, culminating in the initial analysis for the 84-hour NAM forecast. This allows the model to utilize late data from previous cycles and spin up a dynamically consistent field across the entire domain over the 12-hour pre-forecast time period. Previously, the initial guess at the 12-hour-prior time came from the previous NAM assimilation cycle, so the NAM was essentially cycling on itself. Now, the atmospheric fields (e.g., not soil moisture and other non-atmospheric conditions) are taken from the GFS, using the final GDAS cycle which includes late-arriving data. Thus, now, the NAM forecast has a direct link to the GFS analysis from 12 hours ago, though it has been modified by 12 hours of NAM assimilation cycling.

Example:
12 UTC 84-hour NAM forecast
----------------------------------
1. 00 UTC first guess from GFS final analysis (includes data arriving later than from 00 UTC GFS run)
2. observations + regional GSI --> NAM analysis valid 00 UTC
3. 3-hour NAM "forecast" 00-03 UTC --> first guess for 03 UTC analysis
4. observations + regional GSI --> NAM analysis valid 03 UTC
5. 3-hour NAM "forecast" 03-06 UTC --> first guess for 06 UTC analysis
6. observations + regional GSI --> NAM analysis valid 06 UTC
7. 3-hour NAM "forecast" 06-09 UTC --> first guess for 09 UTC analysis
8. observations + regional GSI --> NAM analysis valid 09 UTC
9. 3-hour NAM "forecast" 09-12 UTC --> first guess for 12 UTC analysis
10. observations + regional GSI --> NAM analysis valid 12 UTC
11. 84 hour NAM forecast from 12 UTC

Stephen

A major upgrade to the RTMA was implemented in December 2008.
Details and a case example illustrating the differences are presented on COMET web pages linked from the Operational Models Matrix. The differences are large and noteworthy - you will notice a significant improvement!
Go to http://www.meted.ucar.edu/nwp/pcu2 then scroll down to the Assimilation System section toward the bottom of the page and click on "Updates" under RTMA.

If you notice anything about the RTMA you would like to discuss, any examples of problems you have seen or something it handled very well, or any questions about the changes, please post here and, if you can, attach an image showing what you are talking about.

Stephen

I should have emphasized the issue of representativeness.
Snowshoe is your peak elevation spot. Is this observation really representative of the area around it on the scale of your NDFD grids? Should the gridded forecast issued by the WFO predict the temperature at Snowshoe where the ski resort is located or should NWS be predicting the temperature in the vicinity, which may be quite different due to the mountainous terrain? Does it make sense to use the Snowshoe observation to verify the gridded forecast?

Even if/when NDFD and RTMA move to 2.5 km grids and the corresponding terrain has closer matches to individual observation spots, there will still be some places with considerable terrain-related temperature variability inside a 2.5 km x 2.5 km area on nearly a daily basis. Under some special conditions such as evening decoupling in calm conditions under clear skies, there can even be amazing micro-scale variability over gently sloping terrain in relatively flat areas, such as at El Reno, Oklahoma which I discuss in the RTMA training you can find on the COMET web page http://www.meted.ucar.edu/nwp/RTMA/ (NOAA employees can get credit for it by taking it through the NWS Learning Center at http://doc.learn.com/noaa/nws )

Stephen

The RTMA terrain at this location is approximately 3600 feet, around 1000 feet lower than the ridge-top observation site. During the afternoon on a day with good mixing, we can expect the downscaling from the 13-km RUC to contribute an adiabatic difference compared to observations over this distance (for instance, if the RUC first guess has the correct potential temperature), which explains most of the difference between the RTMA temperature and the observed temperature in this case.

The RTMA will soon be experimenting (not in the operational version you receive, yet) with a terrain that has sharper relief and happens to fit observation elevations closer in complex terrain. There are also plans (contingent on resources being available, etc.) to move to a 2.5-km product which will use a 2.5-km terrain, which fits the observation elevations even better. I have not checked how those compare specifically for Snowshoe, WV, but in general this is an issue common in the western states which will be helped by these other terrain fields. There will still be a lot of places with terrain differences of around 300 feet, but these newer terrain fields greatly reduce the number of stations with very large differences such as at Snowshoe.

Stephen Jascourt

The RTMA has had a lot of difficulty with cold air trapped in valleys and, in general, with situations containing a strong temperature inversion. The downscaling from the 13-km RUC to the 5-km RTMA first guess was changed at 12 UTC on January 8, 2008 to help with this problem. This change helps, though it does not eliminate the problem.

Where the downscaled terrain is below the 13-km RUC terrain, the temperature is now allowed to be up to 10 deg C colder than the 13-km RUC 2-m temperature, based on multiplying the elevation difference by the low-level lapse rate, though it is still not allowed to be colder than the 13-km RUC 2-m dewpoint. Previously, the lapse rate was truncated at isothermal, preventing the downscaled temperature at lower elevation from being colder than the RUC 2-m temperature. However, the dewpoint is still not allowed to be less than the 13-km RUC 2-m dewpoint.

When the downscaled terrain is above the 13-km RUC terrain, the temperature is now allowed to be warmer than the 13-km RUC 2-m temperature, based on the colder of: 1) multiplying the elevation difference by the low-level lapse rate, or, 2) the free atmosphere temperature difference in the RUC between the downscaled terrain and the lowest RUC model layer. Previously, procedure (2) was used except the result was truncated to not allow a warmer temperature than the 13-km RUC 2-m temperature. Also, the dewpoint calculation was changed so that the downscaled dewpoint now uses the 2-meter mixing ratio from the 13-km RUC based on the pressure at the downscaled terrain height.

Stephen

NCEP is in the process of trying to set up a system through which WFOs can supply a list of bad stations they identified in their local area. Meanwhile, you can send me a list (Stephen.Jascourt@noaa.gov) and I'll pass it along. Please indicate whether the temperature, dewpoint, or wind is bad (or which combination) and whether this is permanent or a seasonal problem (maybe the dewpoint bias is big only when the dewpoints are high, etc.) or whether the station is scheduled for repair, etc.
Stephen

Hi Jeff,
Thanks for calling this to everyone's attention.

NAM 2-meter dewpoints have had an extreme high bias over the Appalachians, more than any regional-size area in the past week, though the biggest problems have been over the mountain forests near the west coast. To reduce this, the minimum canopy resistance was increased for evergreen needle leaf forest and mixed forest, resulting in less evapotranspiration in areas dominated by those vegetation types. This has been running experimentally and will be implemented (with no other changes to the model) in a few weeks. The result has been a large reduction in the dewpoints in northwest California on up the coastal ranges (though there still is some problem) but not much relief for the Appalachians.

The problem in your area (West Virginia) is occurring when turbulence shuts down and water from evaporation gets trapped in the lowest model layer. This is not happening uniformly - it is mostly happening after rainfall. For instance, on the 6 UTC run you cite, the 1-hour precipitation at CRW ending at 21 UTC is 0.26", all from the convective parameterization. *AFTER* the precipitation, the surface evaporation cooled the lowest layer (not by much, as the sun was out, so much of the energy for evaporation came from the sun rather than cooling the surface, but it cooled a little), stopping turbulent mixing and throwing a lot of water quickly into the model's bottom layer, which is only 4 mb thick. If you look up just a few layers, say 15 mb above the surface, you will find the much drier dewpoints are still present. The high dewpoints at the bottom are not occurring at other locations where the model did not generate precipitation.

As for CAPE, you need to use a mixed parcel. In AWIPS, the NAM offers the "best" CAPE - most unstable among 6 parcels that are averaged over 30-mb depths in the lowest 180 mb of the column. This field is showing a max of around 1500 J/kg in your area at 21 UTC while the surface CAPE using the bad surface dewpoint is showing nearly 4000. Both increase by 00 UTC, and even the "best" CAPE is probably still be too high - instead, something like SPC's MLCAPE should be used though unfortunately it is not sent over the SBN.

Stephen

This forum is for discussing the Real-Time Mesoscale Analysis, an analysis of surface sensible weather elements made available roughly 40 minutes past each hour. Unique to the RTMA is an estimate of the expected magnitude of the analysis error, providing a sense of uncertainty.

This forum is for discussing how well or poorly the analyses compares with various observations or conceptual sense of observed phenemona, how the analysis is constructed, how it can be used in preparing short term forecasts or in verifying previous forecasts, how it can be used in assessing the current or recent forecasts, how or whether the analysis error size estimate could be used to identify areas where forecasters should focus attention or avoid spending a lot of effort fine-tuning a forecast, and any other such topics as you would like to discuss. Please don't wait for me to post something - this is your discussion forum and what you see may be of interest to others, and others may have useful feedback to help you.

Some background on RTMA:

The RTMA analyses are presently all univariate (no variable influences another). Temperature, dew point, and wind analyses are based on a 2-d var Gridpoint Statistical Analysis (GSI) on the National Digital Forecast Database (NDFD) grid using a first guess from the RUC interpolated down to the NDFD resolution and terrain, along with mesonet observations from all types of mesonets across the country. The hourly precipitation estimates come from the multisensor Stage II analyses (not quality-controlled by the River Forecast Centers - quality controlled data is available too late for RTMA) mosaic produced at NCEP remapped from the HRAP grid, and the sky cover estimate comes from the NESDIS Effective Cloud Amount, which is based on a GOES field of view (larger than NDFD grid box but smaller than a human's visible celestial dome in clear, unobstructed situations) and cannot distinguish between translucent overcast and opaque scattered to broken clouds.

NCEP has been producing the RTMA and sending it over the SBN for many months, and AWIPS has been able to ingest it since OB7.2. However, this original version has too little correction to the RUC first guess. After significant tuning to the GSI and some other adjustments, an improved version is scheduled for implementation tentatively on June 26, 2007. The precipitation and sky cover products are not affected by this change. Concomitantly, COMET will be releasing training on how the RTMA is produced and examples describing the capabilities and limitations of this new version. A link to that training will be posted on this forum.

Even with the June 2007 implementation, a variety of significant changes will still be under development, with one focus on observation quality control and another on allowing the state of the atmosphere itself (e.g., isentropes and streamlines) to affect the shape of an observation's influence on the analysis. The latter would be easy to accomplish if calculation time were unlimited but is a challenge to make fast enough for the short-fuse RTMA.

The RTMA is presently available only over the CONUS. Plans are to implement it over Alaska later in 2007 and subsequently for Hawaii, Guam, and Puerto Rico. The first guess for these will have to be based on the NAM, except GFS for Guam, because these are outside of the RUC domain.

It turns out that the horizontal diffusion change referred to in my last post did not go in yet and is now scheduled for Sep 5. The NCEP parallel experiments which include this change show many shortwave troughs pass more quickly through a long-wave trough axis and move downstream faster. This should somewhat reduce (not eliminate) the problem of troughs typically being too slow and too deep in the NAM-WRF. I do not know why this problem has been persistent in the NAM-WRF and why the modest increase in horizontal diffusion helps to reduce this problem.

Additionally on Sep 5, the SST analysis used in NAM is reverting from the new 1/12th-deg lat-lon product to the 1/2-deg lat-lon product which was used when the Eta model was running in the NAM slot. Large SST problems on the north and west shores of Alaska have adversely impacted the forecast and in the past some isolated excessively warm SST spots in the Gulf Stream were noticed. These artifacts appear to be a result of some problem in the physical retrieval algorithm applied to the AVHRR satellite data. The 1/2-deg product used an older retrieval method. The new method was believed to be superior but apparently it produces bad results under particular circumstances, perhaps related to handling of sea ice in the retrieval (for the problems near Alaska).

Stephen

The Aug 15 changes had a coding problem which caused the increase in horizontal diffusion to be overwritten with the previous, smaller value when the model runs. All of the other changes took effect on Aug 15. The horizontal diffusion increase, after recoding, should actually go in on Aug 22. Horizontal diffusion will still be much less than it was in the Eta model, though a bit more than double what has been running in WRF-NMM.

Therefore, the example Jeff showed does not include the increased horizontal diffusion. Parallel tests now show that it will still produce these spurious events as well as real events, though typically somewhat weaker than without the diffusion increase, and a test run for hurricane Dennis from last year (2005) still develops an intense storm with an eye and eyewall, though those features are sharper when the smaller (current) diffusion is used than when the larger (post-Aug-22) diffusion is used.

As for what is causing these events, it does seem to be the BMJ convective parameterization. Most or in some cases nearly all of the precipitation during the deepening phase in the forecast is from the convective parameterization. Two factors absent in BMJ may be crucial: 1) vertical fluxes of horizontal momentum, and 2) cool convective downdrafts

The GFS decreased its false alarm rate when momentum mixing was added to its SAS convective parameterization, and a few (small sample size) tests in the currently under-development Hurricane WRF also show improvement in reducing false alarms without squashing real storms when momentum mixing is in the SAS convective parameterization compared to when it is not included. WRF doesn't presently allow the convective parameterization to pass along momentum fluxes (being worked on), so this you can't easily test in the workstation WRF at your forecast office.

BMJ does not place cool downdraft air in the subcloud layer (as distinguished, for instance, from SAS and Kain-Fritsch which do). Studies of tropical cyclone genesis indicate that convective downdrafts of much lower equivalent potential temperature than the ambient boundary layer air disrupt surface development, and observations generally show deep moisture up to at least the mid-troposphere in the area where hurricane formation occurs (sometimes not extending much beyond the vortex, but at least encompassing it). This you can test in your workstation WRF by running with KF or SAS. However, it may require a big domain because you'll need to include the area where tropical cyclogenesis is occurring in the operational run, and it may take too long for your local computer to run over such a large domain.

Stephen

The changes mentioned in the previous message are scheduled for implementation Tues Aug 15. Additionally, one more change will also be implemented at the same time: surface (2-meter) temperature observations from land stations, which have not been included in the GSI analysis, will be included in the GSI 3d-var analysis, though with vertical influence confined to be very shallow, primarily affecting the lowest model layer.

Stephen