22 April 2015

What's the difference between GFS and FNL?

[Note: An older version of this information is available under the Documentation Tab of both ds083.2 FNL and ds084.1 GFS as FNLvGFS.pdf.  This page is maintained and should be considered the definitive version.]

GFS stands for Global Forecast System, the current NOAA NCEP global numerical weather prediction system.

FNL refers to the "Final" analysis, though it is now mainly referred to as the GDAS analysis, a part of the Global Data Assimilation System.

FNL and GFS are related, yet different products from the same data assimilation and forecast system. They share the same underlying model and data assimilation techniques. They contain the same data sources--but there is a subtle difference in the amount of "real" data assimilated into the initial conditions for GFS and FNL.

It takes time to run a global NWP model, even with NCEP's vast computational resources. Thus, they need to start GFS early enough to get a forecast instead of a hindcast.

FNL is the final analysis, delayed a bit from GFS so that they can include all of the available observational data. Typically, FNL ingests about 10% more observations than GFS. Even with the late start, it is still available in time so that the 6-hour FNL-based forecast can be used as the background field for the next GFS data assimilation cycle.

For instance, if you want balloon data from 00Z, you have to wait for the balloon to rise through the atmosphere. That can be as much as 90 minutes for the large size balloons capable of reaching the stratosphere. Then the balloon data needs to be relayed from locales around the world to NCEP in Maryland, USA.

If you want satellite data for 00Z, you have to wait for the satellite to go over a ground station so it can download the data. The ground station then relays the data via land or undersea cable (if it has one) or hopscotches it through another (communication) satellite.

NCEP's Current Status table shows the actual start and end times for each analysis/forecast cycle initiated (based on a 30-day running average).

The top of the page provides useful links. The first table refers to the 00 UTC cycle. Note that the North American Mesoscale (NAM) data dump and prep occurs just 75 minutes, 1:15 after the analysis time.  This is possible because of the fast data links available over North America.  The global GFS data dump and prep starts at 2:47 due to the longer data latency for more remote areas.

Data prep includes not just translation of the observations into the ingest data format (BUFR), but also quality control (QC) of the observations.  Observations that are suspect will be given a lower certainty or thrown out altogether.

An analysis is a gridded representation of the atmospheric state that incorporates the observations/measurements in a physically and statistically meaningful way.  They take into account the certainty in each observation.  An analysis begins with a "background" or first guess gridded representation of the atmosphere and then adds the observations to nudge it towards the "true" state of the atmosphere.  The ingest process is repeated with the updated gridded fields over all the observations until a self-consistent grid (atmospheric state) is reached.

Notice that the GFS analysis takes about 20 minutes, about the same amount of time to run the forecast model for 48 hours.

Initiation and end times for the 00 UTC cycle of NAM and GFS.  Click to make bigger.
Scroll further down for the GDAS and Ensemble information.  The GFS "deterministic" model with the best-fit (to the obs) analysis starts at about 03:16 while the "ensemble" of GFS runs with slightly perturbed initial conditions (analysis) starts at about 03:34.

NCEP waits another three hours until 05:50 for more observations to trickle in to repeat the analysis. The second, "final", FNL analysis contains about 10% more observations than the GFS analysis--although the exact percentage varies from day to day. The extra observations ensures the most complete description of the true atmospheric state possible within the operational constraints*.
Initiation and end times for the 00 UTC cycle of GDAS/FNL and GFS Ensemble.  Click to make bigger.
The "final" analysis, FNL, is also called GDAS, because it is an integral part of the Global Data Assimilation System.  The GDAS forecast, initialized off the 00 UTC GDAS analysis, completes around 06:36 UTC.  The forecast valid at 06:00 UTC is then ready for use as the "background" field for the next, 06 UTC, data ingest cycle.   (The 06 UTC NAM cycle kicks off at 7:15 UTC.)

Forecasts contain the same parameters as analysis AND accumulated forecast variables such as precipitation.  Analysis files, if they contain the forecast-only fields, set them to zero.  For instance, both FNL and GFS files obtained directly from NCEP, contain the CRAIN variable.  But, in the FNL file, all of the CRAIN fields will be zero.

In summary, the FNL analysis incorporates ~10% more data than the GFS by waiting for ~3 hours longer for the data stragglers.  GFS forecasts contain accumulated fields, such as precipitation, that are missing or set to zero in the analysis fields.

As an archive, the RDA is interested in offering the most realistic atmospheric analysis, not the timeliest one. Users shouldn't come to an archive for timely data--they should come for the most complete and accurate data possible. We only archive the analysis FNL file--those that end in _00. The forecast ones end in _HH for the number of hours from model initiation.

"Real-time" users can get the FNL analysis grids directly from NCEP's NOMADS server at, http://nomads.ncep.noaa.gov/pub/data/nccf/com/gfs/prod/gdas.YYYYMMDD/gdas1.tHHz.pgrb2.1p00.f000.

Replace YYYY MM DD HH with Year Month Day and Hour of the cycle of interest.

We rename the GDAS/FNL files to fnl_YYYYMMDD_HH_00.grib2 and also strip out the nonsensical forecast variables set to zero in the analysis files.

The NCAR RDA also archives GFS forecasts (ds084.1) for users who need the forecast fields.

* Reanalyses, which do not have operational time constraints, may contain even more observations.


ds083.2 FNL
ds084.1 GFS
Analysis, forecast, reanalysis--what's the difference?

Acronym Guide:

NOAA = National Oceanic and Atmospheric Administration
NCEP = National Centers for Environmental Prediction
FNL = Final Analysis aka GDAS
GFS = Global Forecast System
GDAS = Global Data Assimilation System
HRRR = High-Resolution Rapid Refresh
NAM = North American Mesoscale
RAP = Rapid Refresh

1 comment:

  1. Hi,

    You said: 'The GDAS forecast, initialized off the 00 UTC GDAS analysis, completes around 06:36 UTC.'. When I look at the NCEP's Current Status table, why the GDAS forecast starts earlier before GDAS analysis finishes? Should it wait for it to finish and then initialize the forecast? Thank you


This section is for people who want to discuss using our data holdings effectively. Moderators will delete irrelevant comments.