Proposal Title: Advancing the Unified Forecast System Verification and Diagnostics Capability for Marine Applications
Principal Investigator: Thomas Nehrkorn (AER)
Co-Investigators:
Thomas Hamill (NOAA/OAR/ESRL/PSD)
ABSTRACT:
Ensemble forecast systems are an integral part of NWS’ program for producing skillful, reliable deterministic and probabilistic forecasts. However, statistical postprocessing is often required to transform the raw ensemble model output into forecast products that are most useful for forecasters and the general public. This is to address known deficiencies of the modeling system or the ensembles, such as model biases, under-dispersive ensembles, and limitations on model resolution, and to condense the vast amount of available information into manageable data volumes. This proposal focuses on the generation of a deterministic quantitative precipitation forecast from an ensemble forecast system, as is needed in the National Blend of Models project of the NWS. While the raw ensemble mean is generally better than any single ensemble member in terms of mean squared errors, predicted precipitation features are often unrealistic: because of differences in position of precipitation features in the individual ensemble members, the ensemble mean areal extent is generally too large, and the forecast extrema are too small.
We propose to implement and further develop an ensemble coalescence postprocessing algorithm to address these shortcomings, leveraging previous work by the investigators:
-
A prototype implementation of ensemble coalescence, in which individual ensemble members’ features are displaced towards an ensemble centroid, using a combination of model precipitation and integrated water vapor fields. The resulting “coalesced” ensemble mean has a precipitation feature with a much more realistic areal extent and maximum magnitude.
-
The feature alignment technique (FAT), a variational technique for determining the displacements needed for ensemble coalescence.
-
Quantile mapping, which is designed to correct the raw ensemble mean by shrinking areas of light precipitation and amplifying precipitation maxima. This requires computation of cumulative probability density functions (CDFs) from a historical sample of forecasts and verifying analyses.
We plan to combine these techniques to arrive at a final product that combines the strengths of each method and overcomes their shortcomings. Technique development will be based on GEFS v12 reforecasts over the continental United States, but the developed algorithm will be generally applicable for different forecast models. Performance evaluation will be performed over a full year, using metrics that include standard scores (equitable threat score and bias) and others determined in coordination with MDL.
