On March 6, 2002, NCEP changed the configuration of the Global Forecast System (GFS) to provide four runs of the Global Spectral Model for forecast projections out to 384 hours. These runs of the GFS are generated during the 0000, 0600, 1200, and 1800 UTC forecast cycles, and are initialized with data available by approximately 2 hours and 45 minutes after the nominal initial time (that is, a data cut-off of 2+45 after 0000, 0600, 1200, and 1800 UTC). Prior to this enhancement, these runs of the Global Spectral Model comprised the AVN. Before the reconfiguration of the GFS, the Global Spectral Model, run during the 0000 UTC forecast cycle with a data cut-off of approximately 6 hours, produced the primary model guidance for projections to 384 hours. This particular configuration of the GFS comprised the MRF. Since the changes made on March 6 provide four extended runs of the GFS, NCEP is terminating the MRF, effective April 24, 2002.
The impact of these changes on the current suite of AVN and MRF MOS guidance is minimal as far as the forecaster is concerned. The bias characteristics of the AVN and MRF model runs used to develop the AVN and MRF MOS equations were essentially identical since the same Global Spectral Model was used in both model runs, but with different initial conditions. The different initial conditions imply, however, that forecast variables from the MRF model run might have different correlations with the observed weather than variables from the AVN model runs. Although the Meteorological Development Laboratory (MDL) developed different equations to generate the AVN and MRF MOS guidance, the output from the GFS can be used in either set of guidance equations.
During a 3-week period in March 2002, MDL tested the effects of the GFS change on the MOS guidance by using the new extended 0000 UTC GFS run initialized at 0245 UTC in the MRF MOS max/min temperature and PoP forecast equations. For max/min temperature, the mean absolute errors (MAE) of the test (or parallel) runs were nearly identical to those of the operational MRF MOS guidance at all projections, although some slight increase in the MAE of the parallel MRF MOS can be seen in the forecast projections of 120 to 180 hours See verification results from http://www.nws.noaa.gov/mdl/synop/gfs/mrfmos.html. For PoP, the Brier scores (or mean square errors) of the parallel and operational guidance were also approximately identical; again, a very slight increase in the Brier score was evident in the test runs for projections of 120 to 144 hours. These minor differences in the operational and parallel MRF MOS guidance packages are not significant. Thus, when the MRF is discontinued, MDL will continue to produce extended- range guidance during the 0000 UTC forecast cycle by using GFS variables in the MRF MOS equations, and the forecaster should see little or no difference in forecast quality. No changes will occur in the AVN or MRF MOS guidance packages. In the short term, at least, the nomenclature of calling the latter package the "MRF MOS" guidance will be retained.
In the long term, however, these modifications to the GFS will require MDL to change the development approach used to generate the short- and extended-range guidance packages. MDL is planning to eliminate the two sets of forecast equations (one for the AVN run and one for the MRF run) used to produce max/min temperature and precipitation guidance during the 0000 UTC cycle. When the next MOS development is completed, only one set of GFS-based equations will be required to produce the guidance for these weather elements. Moreover, after a suitable sample of model data is gathered, MDL plans to develop an extended guidance package for the 1200 UTC forecast cycle. Other changes will include the extension of the short-range guidance to projections of 84 hours, and the increase in temporal resolution of the extended- range guidance. Of course, the terminology used to describe these MOS guidance packages will also evolve to reflect the short- and extended-range nature of the guidance, rather than emphasize the differences in model runs.