About learning models...

Storm2k's very own reference library of links, terminology and basic storm information.

Moderator: S2k Moderators

Message
Author
User avatar
ALhurricane
Professional-Met
Professional-Met
Posts: 452
Joined: Wed Jan 08, 2003 12:46 pm
Location: Daphne, AL

About learning models...

#1 Postby ALhurricane » Mon Jul 28, 2003 9:43 pm

I hate to start a new thread on this when I replied already to another one, but I feel led to do so.

I applaud everybody here that wants to learn more about models. Personally, I am about to graduate the Univ. of South Alabama with my meteorology degree. I've already finished all my met classes (just a few junk classes to get through). First hand experience has taught me that the best way to learn about models is to first learn the fundamentals and background of models.

I would not advise trying to take one model and learn how to read it. If you started out by understanding the fundamentals of Numerical Weather Prediction, you would give yourself a much better foundation on which you can expand your knowledge. As I posted in another thread, here is a link in which I think is very beneficial.

http://meted.ucar.edu/topics_nwp.php

The modules there will help you better understand the basics behind these complex models. It will also give you a better appreciation of their limitations. My professors made some of these modules a mandatory read and I am honestly glad he did.

Best of luck!
0 likes   

User avatar
wxman57
Moderator-Pro Met
Moderator-Pro Met
Posts: 22473
Age: 66
Joined: Sat Jun 21, 2003 8:06 pm
Location: Houston, TX (southwest)

Agreed

#2 Postby wxman57 » Tue Jul 29, 2003 6:54 am

I agree, ALhurricane. One needs to learn the basicis of the models first. That's a good web site. There are also a number of Comet courses on modeling. There's a good PPT on the NHC ftp site in Pasch's folder that discusses hurricane modeling, too.

I think maybe that as far as hurricanes are concerned, it's important to note the vast differences in the <b>types</b> of models available. For example:

statistical/climatological: CLIPER, NHC98

statistical/dynamical: BAMS, BAMM, BAMD

dynamical: GFDL, GFS, UKMET, ETA, ECMWF, NOGAPS

As I said, there are some VERY GOOD PowerPoints on the NHC web site relating to hurricane modeling in Pashc's folder. Look here:

ftp://ftp.nhc.noaa.gov/pub/users/pasch/
0 likes   

Guest

#3 Postby Guest » Tue Jul 29, 2003 7:58 am

Alhurricane and wxman - thanks for the insight. I just thought it would be easier to learn one model then try learning all of them - but as you stated which is very logical - to learn the basis. That is what I plan on doing.

Thanks for the links - I'll keep ya'll update on my progress and I'm sure I'll have some questions.

Patricia
0 likes   

Johnny
Category 5
Category 5
Posts: 1428
Joined: Thu Jul 10, 2003 1:35 pm
Location: No Snow For You, Texas

#4 Postby Johnny » Tue Jul 29, 2003 9:55 am

ALhurricane, thanks for taking the time to give us that information. That is exactly what I was looking for. Again, thanks. Like Ticka said, I'm sure I'll will also have questions down the line so stick around. lol
0 likes   

meteorologyman
Category 2
Category 2
Posts: 541
Joined: Wed Mar 15, 2006 6:48 pm
Location: Florida, Kissimmee/St.CLoud

#5 Postby meteorologyman » Mon Apr 09, 2007 9:10 pm

I thought this could be useful for people today whom don't know about this
0 likes   

User avatar
Downdraft
S2K Supporter
S2K Supporter
Posts: 906
Joined: Wed Oct 09, 2002 8:45 pm
Location: Sanford, Florida
Contact:

#6 Postby Downdraft » Tue Apr 10, 2007 10:05 am

Some other things you need to know about the models. Is the system developing or mature? What latitude did the storm form at? Some models don't handle late season storms very well since they can't handle fronts and troughs. There are deep tropics models and mid-latitude ones. You really can't just learn one model. Lastly, those who live by the models die by them too. Maybe I'm to old fashioned but some old fashion weather smarts works too.
0 likes   

MWatkins
S2K Supporter
S2K Supporter
Posts: 2574
Joined: Sat Oct 12, 2002 7:51 pm
Location: SE Florida
Contact:

#7 Postby MWatkins » Tue Apr 10, 2007 12:56 pm

Excellent thread, Jason...

For whatever this is worth, there a few extra caveats that I would throw into this thread...if they apply...

1. Any One Run of Any One Model is Meaningless by Itself So many times, people interpret all of the models, or a particular model, as an expected outcome. In other words, and we will hear this a lot this year I am sure, people will look at a particular run of any model and say something like “oh no, the GFS has this heading right toward me”, without once stopping to consider the model may have had a series 15 solutions 600 miles south of that point.

There may very well be a story being told by that particular solution, but that story isn’t that the hurricane is heading to point A on the map.

When you see something like that, you should ask yourself a series of questions like: What are the other models doing? Is there something in the model that could cause a bias? Watch the next few runs of the model…if they continue to show the shift, then this may be important. But any one run…all by itself and out of the context of the other models, and other runs of the same model, is not terribly useful.

2. Be Suspicious of Big Changes in the 6Z or 18Z Guidance Remember that the global models almost never get new data for the 6Z and 18Z runs, which means these models are rerun using 12 hour old data. This is also why we almost never see huge changes to the 11AM and 11PM EDT forecast tracks from the NHC if the models start to shift…they almost always wait until new data gets in at the 0Z and 12Z runs. Don’t get too excited about a shift until it’s confirmed by the full runs of the models.

3. The GFDL, BAM and LBAR Models are Run Against a GFS Background Any big errors in the GFS will undoubtedly show up in the guidance from these models.

4. Timing is Everything Do you know the difference between the early and late models? Surprisingly, some trained meteorologists don’t. For example, there has been a rumor floating around for YEARS that the NHC holds back the GFDL track data. I have seen it on university websites…but it isn’t true. The GFDL is a late model. It waits for the GFS to finish so it can use the latest output for the track solution…which is why the 12Z GFDL comes out about an hour after the GFS finishes it’s run. LBAR, BAM, NHC98 etc are early models. They run against old backgrounds.

Let’s say it’s 9AM EDT on August 15th, and hurricane Bart is milling around in the Atlantic. You look on the website and someone has posted the latest run of the NHC models. But do you know the timing of the data? Well the BAM’s, LBAR and NHC98 run against the 06Z GFS…not exactly the most reliable data for the “latest models”…huh? Those models are using, essentially, 13 hour old data extrapolated up to the position of features 12 hours ago. See how we got here?

a. 0Z GFS Runs with data as of 8PM EDT last night – Ready at 12:20AM
b. 06 GFS Runs with data as of 8PM EDT last night – Moved up 6 hours 6:20AM
c. 12Z Early Guidance…running off of 13 hour old data

The GFDL, on the other hand, waits for the 12Z run of the globals to finish (a little after noon EDT) before running…and finishes about 1:30PM PM EDT and is available for the 5PM EDT forecast package.

5. Consensus Wins…Almost Always The group of models together (the GFS, GFDL, NOGAPS and UKMET) almost always win over time. This is why the NHC leans heavily on forming an average of these 4 models…this concensus is mostly why track errors have come down so far in the last 10 years…

I am sure there is more to add to this list…but these are the things that popped into my head. Not sure if I am getting away from the original intent of the thread or not…

MW
0 likes   
Updating on the twitter now: http://www.twitter.com/@watkinstrack

wjs3
S2K Supporter
S2K Supporter
Posts: 633
Joined: Sat Aug 14, 2004 7:57 am

#8 Postby wjs3 » Tue Apr 10, 2007 3:20 pm

Hi:

Does anyone want to add any insight on ensembles? How often do you guys look at, for instance, the GFS ensembles? Do you use other ensembles in your forecasting?

Also, I'd love to hear any thoughts on the SHIPS model for intensity forecasting (including how to interpret raw output...still trying to learn how).

I think another point (you all will correct me if I am wrong) is that intensity models are only as good as the track models that drive them. In other words, if track models put a cyclone in the middle of the Gulf of Mexico on day two, and the cyclone is actually, say, over Cuba (Ernesto, anyone), then the intensity models are hopelessly off, right????

Finally, can anyone comment on changes to the SHIPS model this year? I understood from a visit to the NHC that changes were underway this year. Truth to the rumor???

WJS3
0 likes   

User avatar
Janie2006
Category 5
Category 5
Posts: 1293
Joined: Mon Sep 18, 2006 3:28 pm
Location: coastal Ms aka home of the hurricanes

#9 Postby Janie2006 » Tue Apr 10, 2007 7:30 pm

I remember that the UKMET got Ivan's track almost exactly right 10 days away from landfall...we were all rather surprised, to say the least. One of those flukes which turn out to verify.

I just want to have a look at the FSU met computers. I haven't had any luck so far, though. Apparently they remain off-limits to most folks.
0 likes   

User avatar
Recurve
Category 5
Category 5
Posts: 1630
Joined: Tue Aug 16, 2005 8:59 pm
Location: St. Petersburg, FL

#10 Postby Recurve » Wed Apr 11, 2007 7:17 pm

For benefit of novices, it might help to mention the labels on track forcasts
Isn't the consensus the NHC uses often labeled CONU or something you might not get easily?

XTRAP: extrapolated track. A straight line from current position based on direction and speed only.

NHC model summary info:

Table 1: NHC Operational Track Guidance Models Model Type Timeliness
AVN Global baroclinic Late
NOGAPS Global baroclinic Late
UKMET Global baroclinic Late
GFDL Limited-area baroclinic Late
GFDI Interpolated GFDL Early
LBAR Limited-area barotropic Early
BAM Trajectory Early
NHC90/NHC91 Statistical Early
CLIPER Statistical Early

FROM NHC/TPC Tropical Cyclone Models

Brief summaries of common models:

LBAR (Limited area sine transform BARotropic)
Track prediction model, uses averaged winds and heights from the Aviation Run of the MRF global model, with an idealized vortex added to the global model analysis.

GFDL (Geophysical Fluid Dynamics Laboratory)
Limited area baroclinic model developed for hurricane prediction.
Has a specialized method for initializing storm circulation, obtained from the Aviation run of the MRF model. The idealized vortex is based upon parameters of the observed storm.

BAM (Beta and Advection) Model follows a trajectory from the Aviation Run of the MRF model for its track forecast, using averaged horizontal wind. BAM model is run with shallow- (BAMS), medium- (BAMM), and deep-averaged (BAMD) winds.

NHC90/NHC91 and CLIPER
Statistical track forecast models.
CLIPER (CLImatology and PERsistence) uses initial latitude and longitude of the storm, components of the storm motion vector, the Julian day and storm intensity.
NHC90 -- more general statistical model, uses output from CLIPER with vertically averaged heights from the Aviation Run of the MRF model.
0 likes   

Derecho
Category 5
Category 5
Posts: 1011
Joined: Sat Mar 15, 2003 3:15 pm

Key Things:

#11 Postby Derecho » Tue Apr 17, 2007 7:45 pm

I think some serious problems in model understanding have been CAUSED by some of the more popular model sites...

Some of which have been posting UNLABLED "Spaghetti" Plots of EVERY "model" - including XTRAP, 10 different members of the GFS ensembles, lots of ancient stuff, CLIPER, etc.

Looking at those is actually worse than not looking at models at all. What happens is you inevitably end up woth a massive scattershot of lines going every which way that is telling you absolutely nothing and causing people to very wrongly claim " the models are all over the place, this could go anywhere."

Frankly, I'm not sure why some models are still even run, like the useless A98E.

The blunt truth is NHC is only USING a pretty limited set of models 99% of the time for any given forecast.

In fact, MOST of the time now, their forecast tracks are essentially identical to the GUNA or CONU average (which is the GFS, UKMET, NOGAPS, and GFDL, and in some cases the GFDN (GFDL with NGP background) for CONU, added together.) That's basically it, with some attention paid to the ECMWF in the long range - one problem is the EC doesn't come out QUITE early enough for them to use.

Unfortunately a great many "model" plot sites aren't even showing GUNA or CONU at all, or if they are, they're showing them as 1 of 50 model plots with no indication to the uneducated user that they're about 1 trillion-billion times more likely to be accurate than, say, A98E or CLIPER.

There's a bit too much emphasis overall in the layman tropical community, on the "NHC" runs of BAMD, BAMM, A98E, and LBAR too, relative to their usefulness, which is extant but limited for BAMD and BAMM, and practically nonexistent for A98E and LBAR.

Other thing is the GGEM or CMC or "Canadian" is a generally awful model most of the time for tropical systems that gets far too much attention.
0 likes   

Derecho
Category 5
Category 5
Posts: 1011
Joined: Sat Mar 15, 2003 3:15 pm

#12 Postby Derecho » Tue Apr 17, 2007 7:46 pm

Recurve wrote:For benefit of novices, it might help to mention the labels on track forcasts
Isn't the consensus the NHC uses often labeled CONU or something you might not get easily?


CONU shows up in the standard text dumps that a program like Stormtrakker will use to make model plots.

And you can use Stormtrakker to make your own GUNA if you'd like.

Key think is to DELETE all the extraneous crappy models that don't matter like CLIPER and whatnot from those datadumps so you're only plotting the 7-8 models that might really matter.
0 likes   

Derecho
Category 5
Category 5
Posts: 1011
Joined: Sat Mar 15, 2003 3:15 pm

#13 Postby Derecho » Tue Apr 17, 2007 7:48 pm

Janie2006 wrote:I just want to have a look at the FSU met computers. I haven't had any luck so far, though. Apparently they remain off-limits to most folks.



I used to really want to to, but as it turns out, you're not really missing much at all.

The FSU Superensemble is no more accurate than GUNA or CONU, it takes too long to come out, and it "cheats" by incorporating previous NHC forecasts into the superensemble.

It's sort of overrated because it's unavaliable and mysterious.
0 likes   

Derecho
Category 5
Category 5
Posts: 1011
Joined: Sat Mar 15, 2003 3:15 pm

#14 Postby Derecho » Tue Apr 17, 2007 7:53 pm

wjs3 wrote:Hi:

Does anyone want to add any insight on ensembles? How often do you guys look at, for instance, the GFS ensembles? Do you use other ensembles in your forecasting?

WJS3


The average of the GFS ensembles (AEMN and AEMI) had a two-year run where it did really well in 2004 and 2005, but it wasn't very good in 2006.

It's one of the few plots I choose to leave on maps when I use Stormtrakker. I'd take CONU or GUNA over it if it disagreed with it.
0 likes   

wjs3
S2K Supporter
S2K Supporter
Posts: 633
Joined: Sat Aug 14, 2004 7:57 am

#15 Postby wjs3 » Wed Apr 18, 2007 8:11 am

Thanks. What about the GFS ensemble spread--not just the mean.

WJS3
0 likes   

User avatar
ALhurricane
Professional-Met
Professional-Met
Posts: 452
Joined: Wed Jan 08, 2003 12:46 pm
Location: Daphne, AL

#16 Postby ALhurricane » Fri May 04, 2007 2:04 pm

Wow, a thread I started 3 1/2 years ago while still in college has come back to life! What Derecho and Mike added since then are all good points. There are truly only a good few models out there that the NWS/NHC even bother looking at. On top of that, it is always dangerous to follow on deterministic model output, especially if it offers a radical solution. Ensemble models and model consesus is becoming more widely used and has great benefits.

As far as the GFS ensemble spread, it is important to see what the individual ensemble members are saying. If the individual member vary greatly, then the overall mean has to be looked upon with more caution.

The bottom line is that the best way to understand and use these models is to get first become educated in the subject of numerical weather prediction.
0 likes   

User avatar
Scott_inVA
Storm2k Forecaster
Storm2k Forecaster
Posts: 1238
Joined: Sat Oct 12, 2002 5:44 pm
Location: Lexington, Virginia
Contact:

Re: Key Things:

#17 Postby Scott_inVA » Tue May 08, 2007 5:02 pm

Derecho wrote:I think some serious problems in model understanding have been CAUSED by some of the more popular model sites...

Some of which have been posting UNLABLED "Spaghetti" Plots of EVERY "model" - including XTRAP, 10 different members of the GFS ensembles, lots of ancient stuff, CLIPER, etc.

Looking at those is actually worse than not looking at models at all. What happens is you inevitably end up woth a massive scattershot of lines going every which way that is telling you absolutely nothing and causing people to very wrongly claim " the models are all over the place, this could go anywhere."

Frankly, I'm not sure why some models are still even run, like the useless A98E.

The blunt truth is NHC is only USING a pretty limited set of models 99% of the time for any given forecast.

In fact, MOST of the time now, their forecast tracks are essentially identical to the GUNA or CONU average (which is the GFS, UKMET, NOGAPS, and GFDL, and in some cases the GFDN (GFDL with NGP background) for CONU, added together.) That's basically it, with some attention paid to the ECMWF in the long range - one problem is the EC doesn't come out QUITE early enough for them to use.

Unfortunately a great many "model" plot sites aren't even showing GUNA or CONU at all, or if they are, they're showing them as 1 of 50 model plots with no indication to the uneducated user that they're about 1 trillion-billion times more likely to be accurate than, say, A98E or CLIPER.

There's a bit too much emphasis overall in the layman tropical community, on the "NHC" runs of BAMD, BAMM, A98E, and LBAR too, relative to their usefulness, which is extant but limited for BAMD and BAMM, and practically nonexistent for A98E and LBAR.

Other thing is the GGEM or CMC or "Canadian" is a generally awful model most of the time for tropical systems that gets far too much attention.


Well, as someone who does run models commercially, I will agree.

Without attempting to do so, I've noticed a trend of mine over the years doing models at WREL.com and now at Mid-Atlantic Wx.com. The closer a TC gets to land, the fewer models I display. Seems contrary to logic (wouldn't you want the "most" when its the "closest"?). No. Initially, I'll run about whatever is available to give a broad...albeit rather crappy...overview of numerical output. Kinda like a "First Call" snowfall map.

As a TC gets closer, I start dumping the trash to focus in on the best available runs and trends. Sure, I also do a "spaghetti" map LBAR, CLIPers, CMC, NAM et.al.), but it is worthless IMO. My primary map will go with the NHC models, GFS, GFDL (and variants), UKie, NOGAPS, sometimes the MM5 when it looks good, and ensembles including: CONU, GUNS and GUNA. These consensus models are far more important and indeed frequently are hugged by TPC/NHC's black line for good reason.

Early on it is challenging to run the GUNS and GUNA but in my view they are important. The EC isn't public soon enough for me to get current data; that problem isn't going to be rectified soon, so it rarely shows up on my maps.

Very early after initialization (such as the time this is written with Invest 90L coming on line) I might have 15+ models but rarely exceed 10 when a storm gets in...unless there's excellent consensus, or one model keeps digging a different direction and I believe it is plausible (GFDN with Katrina is an excellent example). The Climo/Statistical models are first to go...except the XTRP because the extrapolation displays a storm's heading.

Bottom line is a model map is a tool...if one sees a model map and then reacts, they are LOOKING...they are not FORECASTING.

Good discussion, still valid after 3+ years :wink:

Scott
Mid-Atlantic WX.com
Lexington, VA
http://www.midatlanticwx.com
0 likes   

User avatar
btangy
Professional-Met
Professional-Met
Posts: 758
Joined: Fri Sep 19, 2003 11:06 pm
Location: Boulder, CO
Contact:

#18 Postby btangy » Tue May 08, 2007 10:57 pm

I'm a big fan of ensembles for assessing uncertainty. All the major models have ensembles, but not all are publicly available. The GFS ensemble (MREF) and ETA/NAM ensemble (SREF) are the ones I commonly use. Canadian GEM ensemble is also available. Unfortunately, the ECMWF ensembles are not. One must be aware that ensembles are run at a lower resolution, and like the multi-model ensemble, the mean usually outperforms any single ensemble member.

As for intensity, there isn't as much guidance. The two we use most often is the SHIPS (a statistical-dynamical model) and the GFDL (dynamical model). The GFDL will soon be replaced by the Hurricane Weather and Research Forecast Model (HWRF), so it'll be interesting to see how it performs. I can attest to the fact that there is much research right now going in to improving and developing additional intensity models, so hopefully in the near future, we'll have more and better guidance to improve intensity forecasts.
0 likes   

User avatar
Janie2006
Category 5
Category 5
Posts: 1293
Joined: Mon Sep 18, 2006 3:28 pm
Location: coastal Ms aka home of the hurricanes

#19 Postby Janie2006 » Wed May 09, 2007 11:53 am

I think discussions of model guidance systems are always of great benefit, so I'm glad to see this one "stickied" in the forum.
0 likes   

User avatar
mempho
Tropical Storm
Tropical Storm
Posts: 226
Joined: Fri Feb 10, 2006 11:08 am
Location: Memphis, TN

#20 Postby mempho » Thu May 24, 2007 4:59 pm

ALhurricane wrote:Wow, a thread I started 3 1/2 years ago while still in college has come back to life! What Derecho and Mike added since then are all good points. There are truly only a good few models out there that the NWS/NHC even bother looking at. On top of that, it is always dangerous to follow on deterministic model output, especially if it offers a radical solution. Ensemble models and model consesus is becoming more widely used and has great benefits.

As far as the GFS ensemble spread, it is important to see what the individual ensemble members are saying. If the individual member vary greatly, then the overall mean has to be looked upon with more caution.

The bottom line is that the best way to understand and use these models is to get first become educated in the subject of numerical weather prediction.


Didn't you live in Memphis last year?
0 likes   


Return to “Tropical Reference Library”

Who is online

Users browsing this forum: No registered users and 15 guests