Skip Navigation Links www.nws.noaa.gov
NOAA logo - Click to go to the NOAA homepage National Weather Service NWS logo - Click to go to the NWS homepage
EMC Logo
Navigation Bar Left Cap Home News Organization
Navigation Bar End Cap

         MISSION / VISION    |    About EMC

ENSEMBLES & POST PROCESSING

RESEARCH & DEVELOPMENT


NCEP GLOBAL ENSEMBLE IMPLEMENTATION NEWS



Upcoming NAEFS implementation
TIGGE-GIFS (NCEP/GEFS metadata)
Effective 12Z 23 February 2010 - NAEFS/GEFS upgrade

1) Change horizontal resolution from T126 to T190 out to 384hrs
2) Use 8th order horizontal diffusion for all forecast resolutions
3) Introduce ESMF (Earth System Modeling Framework) – Version 3.1.0rp2 for GEFS
4) Add stochastic perturbation scheme to account for random model errors
5) Add new variables (28 more) to pgrba files for NAEFS data exchange
Effective 12Z 4 December 2007 - NAEFS upgrade

1) GFS bias correction
2) Combination of GFS & GEFS forecasts
3) Probabilistic NAEFS forecasts
4) Downscaled NAEFS forecasts
Effective 12Z 27 March 2007 - NAEFS/GEFS upgrade

1) Ensemble size increased from 14 to 20 members
Effective 12Z 30 May 2006 - GEFS Upgrade and NAEFS first Implementation

    GEFS upgrade:
1) increasing ensemble size from 10 to 14 members
2) adding ensemble control for 0600 GMT, 1200 GMT and 1800 GMT
3) introducing ET to breeding method
    NAEFS implementation:
1) bias corrected forecast
2) ensemble weights
3) forecast anomalies
Effective 12Z 16 August 2005 - Ensemble Upgrade
1) Increased resolution of all members to T126 between 180 hrs and 384 hrs (16 days)
2) Changed initial pertubations from 24 hr breeding cycle to 6 hr cycle
3) Added perturbed tropical storm vortex relocation
Effective 12Z 4 May 2004 -
Bias-corrected QPF and PQPF forecasts
Effective 9 March 2004 -
Ensembles available four times daily with T126 resolution out to 180 hrs
Effective 12Z 29 April 2003 -
Mask Rescaling
Effective 12Z 11 January 2001 -
Increased (T126) resolution for all members until 84 hrs
Effective 20 December 2000 at 12Z -
Change in regional rescaling procedure
Effective 12Z 10 May -
Increased membership at 12Z (10 members)

Effective 12Z 27 June 2000 -
Increased resolution for all members until 60 hrs
Pre-implementation write-up (TPB)  -  Post-implementation results
The NCEP global ensemble forecasting system (Draft TPB)
Effective 12Z April 6, 1999
Increase in initial perturbation amplitude size
Effective 07 December 1998 at 12Z
Change in regional rescaling procedure for setting initial perturbation amplitudes
Effective 06 May 1998 at 12Z
New seasonally varying analysis uncertainty estimates introduced into regional rescaling procedure
Effective March 1997
Ensemble forecast data are available on the OSO server
Effective February 11 1997
Ensemble precipitation forecast data now available
For info on how to access ensemble GRIB files via anonymous ftp, and pull ensemble member identification information from the PDS section of the ensemble GRIB files click here .
Effective 12Z December 20 2000
Change in regional rescaling procedure
A bug has been fixed in the regional rescaling procedure of the initial perturbation generation algorithm. The bug was introduced in June 2000 when the resolution of the ensemble forecasts was increased. At one place in the code, where the winds are extracted from the vorticity for use in the kinetic energy scaling, the list of latitudes to use for cosine weighting was not updated to reflect the change from T62 to T126. The effect of this error was to introduce a smoothly varying error in scaling the kinetic energy in the Northern Hemisphere and a randomly varying error in the Southern Hemisphere. The actual rescaling process is limited using a geographic mask, so the perturbations were not allowed to have large or "spike" errors, but we were introducing smooth latitudinal variations in the amplitude of the perturbations in the NH and noisy variations in the SH. This error was corrected starting with the 12Z cycle on 20 Dec 2000.
Effective 12Z April 6 1999
Increase in perturbation amplitude size
With this implementation, the size of the initial ensemble perturbations were globally increased by a factor of 1.75. This does not directly affect the structure or geographical distribution of the bred initial perturbations.
For the NCEP global ensemble, the size of the initial perturbations averaged over the northern hemisphere extratropics was formerly set to be between 8-9 m in terms of 500 hPa height RMS spread around the control analysis. This value was based on estimates from Savijarvi (1995, MWR, 212-221), who provided analysis uncertainty estimates around 8 m (see his Table 1). There are inidications, however, that this level of perturbation amplitude may not be sufficient:

1) There is a deficit in terms of ensemble spread when compared to errors in the ensemble mean (or control) forecast at short ranges. For example, the 2-day ensemble mean error is 80% larger than the corresponding ensemble spread. "Talagrand" histograms show that at the same lead time 25-30 % of the cases the ensemble misses (does not encompass) the verifying analysis (in excess of what is expected due to limited ensemble size). This also suggests a lack of adequate spread at short lead times.
2) A comparison of ECMWF and NCEP reanalysis fields for January-February for the years 1988-1994 revealed that the average rms difference between the two reanalyses over the NH extratropics is around 21.6 m. Assuming that these two analyses are random samples of the true state, the expected rms error in the analyses is around (21.6/sqrt(2)=)15.3 m.
3) Differences between observations and short-range forecasts or analysis fields also suggest that the error in these fields may be larger than 8-9 m for the 500 hPa height. For example, in February 1999, the 6-hour forecasts (FG) were 15 m away from the radiosonde measurements (OBS) over the northern hemisphere extratropics in an rms sense:
FGERR=FG-OBS=15

Let us assume that the instrumental error affecting the observations (OBSERR) is around 10 m. One can write the square of the error in the first guess (FGERR) as:

(FG-OBS)**2=(TRUTH+FGERR-(TRUTH+OBSERR))**2=15**2

Assuming that OBSERR and FGERR are uncorrelated: FGERR=11.2

Using independent data we estimate that the rms error in the analized fields is on the order of 6% smaller than in the first guess forecasts: ANERR=10.6

Given the assumptions, this is an estimate for the error (with respect to the true state of the atmosphere) in the analyses in the vicinity of existing radiosonde sites. We further assume that the analysis uncertainty is substantially (50-100%larger over areas void of radiosondes, and that these areas constitute roughly half of the NH extratropics. The range of RMS uncertainty for the whole NH extratropics then is: ANERR(NH) ~ 10.6 x 1.25 - 10.6 x 1.5, i.e., it is between 13 and 16 m.

Based on the above considerations a decision was made to increase the initial perturbation size over the NH extratropics from around 8.5 m, by a factor of 1.75, to around 15.5 m. This new perturbation level is consistent with analysis/short-range forecast error estimates discussed above.

The increase in perturbation amplitude size is applied globally and should be beneficial to users of forecasts over the NH extratropics and the tropics. The factor of 1.75 increase may have neutral or somewhat negative effect over the SH extratropics where the perturbation amplitudes appeared to be large enough before the implementation. The current perturbation rescaling procedure, that uses monthly varying geographically dependent estimates of analysis uncertainty (based on a comparison between two independent analysis cycles), however, did not allow for controlling the perturbation amplitudes separately over the two hemispheres. Our plans call for the implementation of an adaptive rescaling procedure that would be based on the distribution of analysis and first guess increments to the actual observed data.

Zoltan Toth, Richard Wobus and Istvan Szunyogh (credits to Eugenia Kalnay, Tim Marchok and Suranjana Saha)
Effective 07 December 1998 at 12Z
Change in regional rescaling procedure for setting initial perturbation amplitudes
A bug has been fixed in the initial ensemble perturbation generation code. Earlier, the pertubation amplitudes were too small at low, and too large at high latitudes, due to an incorrect computation of the norm (kinetic energy) used in the regional rescaling procedure (Fig. 1: Smoothed average rotational kinetic energy of initial pertubations around the 500 hPa level for the period February 1-28, 1998, with the bug). With the correction, the overall perturbation amplitudes over high latitudes decreased, while over low latitudes increased; the midlatitude amplitudes remained unchanged (Fig.2: Same as Fig. 1 except with corrected code). The geographical distribution of the resulting new initial perturbations is consistent with that of estimated analysis uncertainty (Fig.3: Geographical "mask" used in the regional rescaling procedure - Smoothed average rms difference in rotational kinetic energy around the 500 hPa level between reanalized and operational NCEP analyses for the period February 1-28, 1995). After the code correction, the global pertubation amplitude was set at such a level that the average size midlatitude pertubation amplitudes, which were consistent with different estimates of analysis uncertainty, remained unchanged.

Zoltan Toth, Istvan Szunyogh and Richard Wobus
Effective 06 May 1998 at 12Z
New seasonally varying analysis uncertainty estimates
On May 6, 1998 at 12Z a change was implemented in the NCEP global ensemble forecasts. The time independent geographical mask (that sets the amplitude to which the 24-hr bred perturbations are rescaled) was replaced by a new mask. This mask is based on the average difference between a series of operational T126 analyses and T62 reanalyses. Monthly average rms difference fields are used (based on one year of data) and are linearly interpolated for each day. This change should provide more realistic initial perturbation amplitudes. In particular, initial perturbation amplitudes will be somewhat reduced at high latitudes whereas somewhat increased at low latitudes.

Zoltan Toth, Gopal Iyengar and Richard Wobus
Effective March 1997
Ensemble data now available on OSO server
Beginning in March, 1997 the NCEP global ensemble data were placed onto the OSO server on a daily basis (in addition to having the data on the nic server). The planned retaining period for the data on the OSO server is 24 hours (though initially this period may be shorter). Currently, the retaining period for the data on the nic server is about 4 days.

Yuejian Zhu and Zoltan Toth
February 11, 1997
Probabilistic Precip GRIB files available via FTP
Beginning Feb. 11, 1997 we provide grib files on our ftp server containing the probability that 24-hr forecast precipitation amounts will exceed certain threshold values. There will be one file per day, based on the latest 17 ensemble members available at 00Z, with a filename like ensppf.97021100. The grib files have a special extension in their headers describing the data. Please consult the appropriate NCEP GRIB documentation for details. For those who cannot (or does not want to) decode the special grib header information, each file contains 133 records. The records contain the following information:

GRIB Record Number Description
1 0-24 hr prob that precip amount > .254 mm
2 0-24 hr prob that precip amount > 1.0 mm 
3 0-24 hr prob that precip amount > 2.54 mm 
4 0-24 hr prob that precip amount > 6.35 mm 
5 0-24 hr prob that precip amount > 12.7 mm 
6 0-24 hr prob that precip amount > 25.4 mm 
7 0-24 hr prob that precip amount > 50.8 mm 
8-14 Same as Records 1-7, except for 12-36 hr fcst interval 
15-21 Same as Records 1-7, except for 24-48 hr fcst interval 
22-28 Same as Records 1-7, except for 36-60 hr fcst interval 
127-133 Same as Records 1-7, except for 216-240 hr fcst interval


The forecast probability is estimated directly from the 17-member global ensemble. At each gridpoint the number of ensemble members having a 24-hour precipitation amount greater than the limit considered is counted (M) and the probability is expressed as 100*(M/17)

Yuejian Zhu and Zoltan Toth