Monday, 12 January 2009: 1:30 PM
What causes NCEP GFS model forecast skill “dropouts”?
Room 130 (Phoenix Convention Center)
Numerical weather prediction (NWP) models use a wide array of conventional and non-conventional observations estimating the state of the Earth's environment. Successful assimilation of observations involves sophisticated algorithms and techniques for quality control (QC) and analysis. Models that embody the physical laws governing the behaviour of the Earth's land surface, oceans, and atmosphere, and computers with the power to run these models rapidly enough to make timely predictions, are also essential elements of an effective environmental analysis and prediction system. NCEP's Global Forecast System (GFS) model forecast cycles every 6 hours generating a background (guess) for the next analysis. The interplay between QC, the assimilation of various observations, the analysis constraints on the incremental changes from a previous forecast, and the character of the model guess as “memory” of past events, is studied for GFS low skill forecasts. To analyze low model forecast skill, we compare the GFS analysis with the European Centre for Medium-range Weather Forecasts (ECMWF) model/analysis. Treating the ECMWF gridded initial conditions as pseudo observations, and using them as sole input into the GFS Grid-point Statistical Interpolation (GSI) analysis which acts as a “grand” interpolator, new initial conditions are generated thereby inheriting ECMWF forecast system characteristics. From these initial conditions, forecasts are made for comparison with NCEP's operational forecasts (control) to detect differences in time, space, and for analysis of QC. We show statistics and construct a climatology of events when skill score “dropouts” occur, and determine if QC problems originate from particular observation types. The goal is to make improvements in both automated QC and the use of observations to alleviate dropouts in model forecast skill. Substituting the ECMWF initial conditions in select areas can show significant GFS forecast skill enhancements for most “dropout” cases and defines an area of model sensitivity. In these sensitive areas QC differences can lead to large impact. We show an example with Quick-Scat observations.