Changes to the 2002 NCEP Operational MRF/AVN
Global Analysis/Forecast System:

Mark Iredell, Hua-Lu Pan and Peter Caplan

National Centers for Environmental Prediction

Global Modeling Branch

W/NP23, World Weather Building,
Washington DC 20233, USA

1. Introduction. On 29 Oct 2002, the followingchanges will be implemented in the MRF/AVN analysis /forecast system (now called the Global Forecast System):

  • Enhanced resolution in the horizontal and vertical
  • Changes in reduction of resolution as the forecast progresses
  • New background error covariances for new vertical resolution
  • Use of additional satellite data
  • Bug fixes

    2. Resolution changes.

  • Initial resolution (out to 84 hours)
  • Current:T170L42
  • New:      T254L64
  • Telescoping during the forecast
  • Current: T170L42, 0-180h; T126L28, 180-364h
  • New:      T254L64, 0-84h;    T170L42,    84-180h; T126L28, 180-384h
  • The following changes will be implemented in the ensembles:
  • Current: T126L28, 0-84h; T62L28 84-384h
  • New:    T126L28, 0-180h; T62L28 180-384h
  • 3. Arrangement of layers.
    The distribution of layers in the vertical is shown in Fig. 1 and in the table below:
    . . . . . .
    comparison of vertical structures
    Layers 64 42 28
    Top (hPa) .266 2 3
    Layers above 100 hPa 24 10 7
    Layers below 850 hPa 13 10 6
    Thickness of lowest layer(m) 20 30 40

    4. Other changes

  • Background error covariance recomputed for new vertical resolution.
  • Satellite data: Increased use of AMSU/A Channel 13, NOAA-16 HIRS and NOAA-14 MSU, Channel 14 (over land).
  • Bug fixes: Various small corrections and changes were made, including
  • Modifications to the 6th order horizontal diffusion for the T254L64 portion
  • Modifications to the spectral transform
  • Allowance for large snow depth over permanent ice
  • Further analysis and data changes added after additional evluation:
  • assimilate METAR surface pressure observations
  • turn off divergence tendency constraint
  • further update of background error
  • thin TOVS data toward polar regions by selection of best data in each box
  • 5.Impact of the changes on model behavior - subjective evaluations from the field.

  • HPC: The consensus at HPC is that the new model showed a small improvement over the old in mass fields and in precipitation. During regimes that produced poor forecasts for the old model, the new suffered similarly.
  • South American/Tropical Desk: Indications were found that the mass fields of the two models were similar, but the new model at times produced precipitation "bombs" in areas with weak synoptic forcing, There did not seem to be any geographic pattern to these occurrences.
  • MPC: Forecasters noted that central pressures in cyclones were usually deeper in the new model. The consensus in this area was that the new model was preferred.
  • TPC: Tropical cyclone track errors out to 72 hours were similar at the two resolutions, but there is some concern about larger track errors seen in the Atlantic beyond 72 hours in the limited number of cases available.
  • AWC: Forecasters at AWC were enthusiastic about the new model, finding improvements in day-one forecasts of clouds, precipitation and icing.
  • SPC: The higher resolution version of the model is found to be "at least as good" as the current model in short-range forecasts of convection and fire weather; otherwise, little difference was seen between forecasts out to 84 hours.
  • MDL: Three months of tests over the U.S. showed no significant changes in MOS between the current and new models.

    6. Model performance statistics vs. analysis - mass and wind fields In evaluations of forecast mass and wind fields against verifying analyses, the new package showed consistent improvement through a period of testing extending from September 2001 through April 2002 The 500-hPa anomaly correlation dieoff curves for the period 10 October 14 April 14, 2002 given in Fig. 2a for the Northern Hemisphere and Fig. 2b for the Southern Hemisphere, show the same general picture. The new model (MRFX, blue)gets better scores on average for all forecast lengths out to 7-8 days, with slightly less skill for days 8-9. This is true in both hemispheres, for all wave number groups shown (an additional one-month test of the "further" analysis changes alluded to above produced comparable results in the Northern Hemisphere and further improvement in the Southern Hemisphere). It is interesting to note that in spite of the decline after day 7, the score for the week-2 average forecast is noticeably better in the new than in the operational(current), as can be seen from Fig. 3 The consistency of performance over time is shown in Fig. 4 in the form of month-by-month averages, new vs current for 5-day forecasts. Another measure of performance, the frequency distribution of scores is shown in Fig. 5 , which indicates that the new model had a lower frequency of poor scores for 5-day forecasts at 500 hPa. The daily 5-day scores (with a 1-2-1 time smoother) show very similar behavior in the two models (Fig. 6). , In the tropics (20S-20N), Fig. 7 shows that there was no significant difference between the old and new models for 3-day wind forecasts at 850 hPa and a slight increase in error at 200 hPa, as verified against analyses.

    7. Model performance statistics vs. analysis - temperature The current model has a cold bias over land in the cold season. The new model shows similar behavior but with a somewhat reduced magnitude of the bias at forecast day 5, as shown in Fig. 8 The reduction in bias continues growing with time out to 16 days.

    8. Verifications against rawinsondes Over a period of five months, the verifications against rawinsonde observations for two-day wind forecasts at 200 hPa smoothed with one-week running averages for the Northern Hemisphere indicate little difference between the models (Fig. 9a) but when broken down by area, the results did reveal small differences over Asia, where the current model was better (Fig. 9b) and over North America, where the new model won. (Fig. 9c)

    8. Precipitation verifications Based on observations over the Continental U.S., the new model yielded much the same threat scores as the current for low and moderate amounts, but somewhat higher for heavy amounts (over 1.5 inches = 38 mm)for forecasts for the 24 hours ending at hour 36. Results only for the month of April are shown (Fig. 10a), but similar results were found in the preceding months. The bias scores in Fig. 10b show that the new model is comparable to the current for small amounts, but wetter at moderate and heavy amounts. In the wettest categories it seems to be considerably closer to a bias of 1.0 than the current model.

    9. Tropical storm tracks Judging from the limited number of cases available, the two models performed at comparable levels in both the Atlantic Fig. 11a, and the Pacific, Fig. 11b, with some questions to be resolved about longer time ranges in the Atlantic.

    10. Summary and future work Our conclusions are that the increase of horizontal and vertical resolution in the new model has improved forecasts slightly at almost all time ranges. Objective evaluations indicate slightly better mass and wind fields but show little effect on the accuracy of tropical storm tracks and precipitation forecasts over the limited time/area where these were evaluated. Subjective evaluations from various users mostly indicate little difference in model behavior, but an overall preference for the perceived better definition of synoptic features in the new model.