Information About Gfs And Its Data Set

[johnholmes]

hi

There have been a nymber of comments over recent weeks about what is in each 'run' with GFS. Where does the information come from. Are the 'teleconnections' part of that data etc. I exchanged e mails with NOAA about this and had their answer back.

Below is the exchange and below that a list of links to see exactly what does go in and a complete list of the way in which the multitude of parameters are dealt with.

This will be placed in the NW Guides section.

I asked

> 1. What do you 'feed' into your model for it to go out to T+384 hours?

> Is it simply the basic data available, surface/upper air etc, every 6

> hours and then run using the basic laws of thermodynamics? Or are other

> factors fed in?

> 2. Once you get beyond what I would call synoptic forecasting, about

> T+144/168, although I know ECMWF are charged with attempting to make it

> synoptic type predictions to T+240, do you factor in items like

> ENSO/AO/NAO? If so how, is there a weighting factor as one gets further

> down the T+line?

I will try to answer your questions. If I don't do them justice, please

let me know.

1. All models get the same input to start them off whether they are run

out to 12 hours or 384 hours. They begin with an analysis that is a

combination of a short-term forecast and all available data at the

time. All of the data is interpolated to the model's grid using data

assimilation techniques. The model uses this as a starting point and

using the basic thermodynamic equations and as sophisticated

parameterization of meteorological processes as the power of the

computer being used will allow, proceeds merrily forward in time,

spitting out forecasts of temperature, wind, pressure, moisture and

whatever other parameters might be programmed in.

2. Nothing particularly special kicks in at day seven and beyond, as far

as the model goes. We know that accuracy can decline past the useful

point once most models reach day seven or eight. In the U.S., once we

are beyond that timeframe, we do not issue forecasts like "the high will

be 73, with a 30% chance of showers in the afternoon." In the eight to

14 day range, we issue categorical forecasts like "temperatures will be

above normal" or "precipitation will be below normal". At this point

in the evolution of forecasting, that is as far as skill of the models

will take us. Here is a link to our forecasts beyond day seven so you

can see the types of forecasts that are possible:

http://www.cpc.ncep.noaa.gov/ We do run a numerical climate prediction

model that runs out to 9 months. But the output of this model is only

one source of what goes into compiling our seasonal forecasts.

As this did not appear to answer my query about the longer term drivers I asked again about ENSO/AO/NAO etc.

this was the reply I got

Specific ENSO, AO, NAO information is not fed into the numerical

models. However, one could say that since each of these indices is a

reflection of the state of the atmosphere and oceans, and the current

state of the atmosphere and oceans (pressure, temperature, winds, etc.)

is known to the models through their initialization, then the models

do know a bit about what the indices are attempting to quantify.

To me that appears to clear up any ambiguity. The ONLY data fed in is the observed data at the start of each run

Below is a comprehensive list of links about what is fed in and how and when for each run.

To me that appears to clear up any ambiguity. The ONLY data fed in is the observed data at the start of each run.

There are also the Ensembles, see below for an explanation of what they are

http://www.netweather.tv/forum/index.php?showtopic=25663

See the link below as to what this is.

http://www.netweather.tv/forum/index.php?showtopic=35985

To me that should make it clear that there is no such thing as, 'GFS defaulting to its Atlantic preference' or something along those lines?

At T+00 the operational run is fed the basic raw data and goes on out to T+384 with NO other input. The software/programme showing what this involves is pretty complex; but that is all that is fed in.

Below are some of the links into NOAA to 'see' what data is routinely fed in and for updates to this.

http://www.emc.ncep.noaa.gov/modelinfo/index.html

http://wwwt.emc.ncep.noaa.gov/gmb/moorthi/gam.html

http://www.nco.ncep.noaa.gov/pmb/products/...z.pgrbanl.shtml

The list above is only a fraction of what you can browse through!

And for anyone still wanting more information then this link MAY help or give you some indication on items mentioned in various model threads

http://www.netweather.tv/forum/index.php?showforum=5

It is a link to the NW Guides with information on most topics.

John Holmes

11 December 2008

[Iceberg]

Thanks John, ECM is slightly different to the GFS is that they can and do force initialise certain events, to my knowledge this is only used for small tropical systems that are too small for the grid point to manage but they know that it can have significant effects downstream in the timeline. I believe ECM is the only model that can do this.

[Methuselah]

Thanks for that, John. It is reassuring to know, rather than merely suspect, that the GFS does NOT have an "Atlantic bias"!

The weather does though, more's the pity.

[Evo]

Unfortunately personal circumstances don't allow me to read or post much any more but during a quick peek this morning saw this topic which is certainly interesting. I'd recommend those interested in this suject to take a look at the Google Books preview of "Atmospheric Modeling, Data Assimilation, and Predictability" by Eugenia Kalnay here:

Google Books

... in particular pages 13 and 14 and note how a short term forecast is used to populate model grid points due to the lack of available observations. There are also some interesting details on the evolution of the modelling process and the problems associated with increased resolution. There's plenty of heavy maths in there too for those inclined!

[J07]

Unfortunately personal circumstances don't allow me to read or post much any more but during a quick peek this morning saw this topic which is certainly interesting. I'd recommend those interested in this suject to take a look at the Google Books preview of "Atmospheric Modeling, Data Assimilation, and Predictability" by Eugenia Kalnay here:

Google Books

... in particular pages 13 and 14 and note how a short term forecast is used to populate model grid points due to the lack of available observations. There are also some interesting details on the evolution of the modelling process and the problems associated with increased resolution. There's plenty of heavy maths in there too for those inclined!

That book's a classic, it even has colour pictures!

I can't get 13 and 14 on my preview and don't have the hard copy around, but assimilation of real data is something that should improve over time. (Warning: terrible rambling coming up ahead.....) From memory, there are various approaches they have used in the past, and one of them is where all observations within T +/- t_c are assumed to have been valid at analysis time T (and where t_c is whatever critical value they chose, eg 3 hours). The problem, naturally, with this is that we see in fast moving systems (and slow moving systems) all the time that the model is always wrong to some extent. A model analysis field is never truly representative of reality, that's why the fax charts are so important. So from this POV, if an assimilation technique assumes that an ob from T-3hrs was taken at T hours, then that could cause problems.

The other thing is the weighting of observations, depending on their distance from a model grid point. I'd be quite interested to find out how exactly they do this in the GFS and IFS models which don't use fixed grids. My assumption is that the observation set is collated to form an observation "wave" in spectral space which is then assimilated into the model. That sounds intuitively correct....but another possibility is that they convert the spectral model output to a fixed grid (which is hideously expensive apparently), assimilate the obs in their natural form, and then convert the whole thing back to spectral again. That sounds really clunky though. If anyone knows how they actually do it, please pipe up!

And then they have to eliminate discontinuities, in that as you move away from an observation, the adjusted model field needs to relax back towards the pre-assimilation model field, otherwise you could end up with something like a discontinuity in the pressure field, which implies an infinite pressure gradient force and hence imminent disaster for the planet!

I think many mesoscale models prefer to use "Newtownian Adjustment" as their scheme, because it assimilates observations at their proper time (hurrah!).

Another cracking benefit of this method is that it incorporates balancing of the fields as a build in safeguard to prevent, for example, spurious gravity waves forming in the early stages of integration, then dissipating sometime later and releasing huge amounts of energy into the model fields and stuffing up absolutely everything.

However, from a philosophical POV, the NA scheme to me may be practical, but it feels artificial and "forced"....because it is. Of course in practical considerations this is meaningless, but it does give us some humbling experiences in how the models must *constantly* be grabbed by the scruff of the neck and dragged back into reality all the bloody time! Models, even at T = 0 are not the truth and never will be, so thank God for people!

I think NA is used in global models too, not sure which ones.

[If Wishes made Weather]

Would it be correct then to state that greater accuracy in models would be served by increased number of accurate observations (higher density of weather reporting stations) being fed into the initial input?

T

[J07]

Would it be correct then to state that greater accuracy in models would be served by increased number of accurate observations (higher density of weather reporting stations) being fed into the initial input?

T

Often yes, if the observations are assimilated properly. If they aren't, then it's possible for there to be a negative impact. A good example is that of satellite derived soundings. Because you cannot set off weather balloons from every spot on the earth, there's huge gaps in vertical temperature profiles. Obviously, these gaps are greatest over the oceans.

If you then use satellites to derive vertical temperature profiles over the oceans, where actual soundings are sparse, the effect is positive on the models' initialisation (and hence forecast fields). But, if you use satellites for this purpose over land, you have to be very careful because they can have a negative impact. This happens because they can disagree with actual soundings. Or perhaps not disagree, but derive much less information (satellites won't pick up on sudden changes in lapse rates, such as inversions, too well), and hence you end up with a large array of "low resolution VTPs", which may not be exactly in tune with the higher resolution VTPs from balloon soundings. So then the modellers have to be very careful in the statistics they use to assimilate satellite soundings when they are close to actual balloon soundings. Because you end up going from one point with high vertical resolution, to another with low vertical resolution, and this could create discontinuities (for example, an inversion may disappear quite suddenly from a balloon point to a satellite point).

They probably approach this by giving less statistical weight to satellite soundings when they are close to a balloon sounding. And then some additional process will have to smooth everything out across the entire horizontal grid to keep the model fields continuous, and yet also agree as much as possible with observations.

[J07]

Often yes, if the observations are assimilated properly. If they aren't, then it's possible for there to be a negative impact. A good example is that of satellite derived soundings. Because you cannot set off weather balloons from every spot on the earth, there's huge gaps in vertical temperature profiles. Obviously, these gaps are greatest over the oceans.

Low quality, but good enough, image from the WMO showing distribution of sounding sites.

From a network of roughly 900 upper-air stations, radiosondes, attached to free-rising balloons, make measurements of pressure, wind velocity, temperature and humidity from just above ground to heights of up to 30km. Over two thirds of the stations make observations at 0000UTC and 1200UTC. Between 100 and 200 stations make observations once per day. In ocean areas, radiosonde observations are taken by about 15 ships, which mainly ply the North Atlantic, fitted with automated shipboard upper-air sounding facilities (ASAP). A subset of upper-ait stations comprise the GCOS Upper-air Network (GUAN). For more details of the upper-air stations see the catalogue of upper air stations.

http://www.wmo.ch/pages/prog/www/OSY/gos-components.html

Edited December 19, 2008 by J07