| Summary: | Numerical weather prediction models, which are discretized approximations to the
physical equations of the atmosphere, are a critical part of weather forecasting. They
can project an observed state of the atmosphere into the future, but forecasts have
many error sources. To counter this, several forecasts made from slightly different
initial states or models can be made over the same domain and time period. This
approach, called ensemble forecasting, can forecast uncertainty, and produce a better
average forecast.
In this study, ensemble forecasts are generated and analyzed to understand the
nature of initial condition (IC) and model error over the North Pacific. A poorlypredicted
storm event (a bust) in Feb. 1999 is a useful case study period. To approximate
different aspects of IC uncertainty, three IC-perturbation methods are used:
(1) ranked perturbations that target coherent structures in the analyses; (2) perturbations
that simulate differences between operational analyses from major forecast
centers; and (3) random perturbations. An ensemble of different model configurations
approximates model uncertainty.
Ensembles are verified several ways to separate model and IC uncertainty, and
evaluate ensemble performance. It is found that during the period surrounding the
bust, IC error is greater than model error. But for one critical forecast, model error
is unusually high while error from ICs is unusually small.
Comparison with rawinsonde observations shows that differences between operational
analyses cannot account for analysis error. Ensembles generated with this
information show that analysis differences contain some spatial information about
analysis uncertainty, but the magnitude is too small. To account for total forecast
error, model uncertainty must be included with IC uncertainty.
Comparing different ensembles reveals that a scaled ranked-perturbation ensembles
has the best characteristics, including perturbation magnitude, uncertainty growth
vs. error growth, spread-error correlation, and shape of the variance spectrum. Its
properties are verified by running an independent case study, and only minor differences
are found.
Contributions to the field of weather prediction include a new ranked-perturbation
method that results in improved short-range ensemble-mean forecasts, an understanding
of the relationship between analysis error and analysis differences, and the confirmation
that model uncertainty is necessary to account for short-range forecast error.
|
|---|
