Global Forecast Dropout Prediction Tool (GFDPT) Future Strategies at NCEP

Poor forecasts or skill score dropouts occur in the National Weather Service (NWS) Global Forecast System (GFS) often when other national center forecasts—for example, the European Centre for Medium-range Weather Forecasting (ECMWF)— do not exhibit a similar loss in skill. The GFDPT collaborative project between JCSDA/NESDIS and NCEP/EMC evaluates model forecasts focusing attention on the assimilation system and real-time observation-quality feedback into the model development cycle. The GFDPT detects regional differences between the NCEP and ECMWF global models operationally and determines if dropouts originate from QC problems in the assimilation, especially the assimilation of radiances. The various components of the GFDPT web based system, for use in NCEP’s Weather Prediction Center (WPC) and Model Evaluation Group (MEG), are the depiction of actionable volumes of conventional and satellite observations that potentially cause the dropout. The forecast-forecast correlations between ECMWF and GFS provide first warning of GFS dropout potential, an extremes code sifts out and displays extreme volumetric integrals of squared differences of geophysical fields of GFS analysis differences compared with ECMWF and other background guess fields, and finally a full diagnosis of actionable volumes is performed using the Gridpoint Statistical Interpolation (GSI) conventional data and satellite radiance data diagnostics files.

Results show that improving the initial condition (IC), by use of ECMWF analysis converted to pseudo observations and placed in the regions that caused the dropout, improves the GFS production 5-day forecast skill. We show evidence that the occurrence of rare poor skill forecast events in the GFS production, and the difference between GFS and ECMWF are attributable to possible contaminated radiance assimilation causing errors and conditioning of the GFS IC. Differences between ECMWF and GFS IC seem to correlate with the largest differences in observed radiance and that from the analysis projected by the radiative transfer model (O-A). The background guess contribution (from model and physics) may also contribute as shown in a case study. We show an Observing System Experiment (OSE)/observation error optimization case study to confirm that forecasts are improved by removing/adjusting observations a priori by improving the radiance quality control. Removing the surface channel radiances from the analysis over dropout source areas caused improvement in the skill. Successful prediction of low skill score events with F-F correlation, extremes between ECMWF and GFS are found, radiance O-A charted per channel regionally and co-located with the locations that are found to be responsible for the dropout can be more widely used for model evaluation.

The GFDPT team initiated a collaboration to test an optimal combination of GFDPT and Ensemble Forecast Sensitivity to Observations (EFSO) and Proactive Quality Control (PQC) developed at University of Maryland at NCEP. This provides an efficient framework for implementation of new observing systems with advanced QC to address a continuous monitoring of the quality of every observation and most importantly, PQC, which promises to identify and remove detrimental observations with a significant accumulated improvement of GDAS, leading to large improvements in the 2-5 day forecasts.