Thursday, 15 January 2004: 2:00 PM
CONSTRUCTION OF AN ENSEMBLE OF FORECASTS USING ADJOINT-DERIVED SENSITIVITIES
The adjoint of a numerical weather prediction (NWP) model can be used to
evaluate the forecast sensitivity gradient: how changes to initial conditions
of an NWP model forecast will affect some function of the model forecast state
(called the response function, R). Given a forecast sensitivity gradient (with
respect to initial conditions) for one particular forecast aspect, and given
some knowledge of the analysis errors, it may be possible to estimate the
likely ranges of values the response function may take. This may be an
effective way of generating an ensemble of forecasts, as it requires a single
integration of both the NWP model and its corresponding adjoint. In addition,
this method allows a forecaster to construct an ensemble which may be tailored
for specific forecast needs (site or regionally specific, as well as a wide
variety of forecast aspects).
In this presentation, an ensemble of short-term forecasts of a particular
forecast aspect are generated using forecast sensitivities of that forecast
aspect with respect to model initial conditions (using the MM5v2 Adjoint
Modeling System). The estimation of initial condition uncertainty is generated
by using the differences in MM5 initial conditions derived from different
analyses from several operational models. The product of the analysis
differences with the sensitivities determines the approximate changes in the
response function at the end of the forecast interval. This calculation allows
bounds on a particular forecast aspect to be established.
A more traditional method of ensemble generation is used to determine the
efficacy in using the previously described methodology, as well as determining
the validity of the tangent linear assumption (the basis for the adjoint
model). A set of non-linear model integrations forward in time using perturbed
initial conditions (generated from the same analysis differences) are being run
for this purpose.
A statistical analysis of the results of several months of running this ensemble
forecast generation procedure is presented and suggestions regarding the
validity of this methodology are discussed.