 |
| Titel |
Applying a fully nonlinear particle filter on a coupled ocean-atmosphere climate model |
| VerfasserIn |
Philip Browne, Peter Jan van Leeuwen, Simon Wilson |
| Konferenz |
EGU General Assembly 2014
|
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 16 (2014) |
| Datensatznummer |
250092342
|
| Publikation (Nr.) |
 EGU/EGU2014-6678.pdf |
|
|
|
|
|
| Zusammenfassung |
| It is a widely held assumption that particle filters are not applicable in high-dimensional
systems due to filter degeneracy, commonly called the curse of dimensionality. This is only
true of naive particle filters, and indeed it has been shown much more advanced methods
perform particularly well on systems of dimension up to 216 - 6.5 x 104. In this talk we will
present results from using the equivalent weights particle filter in twin experiments with the
global climate model HadCM3.
These experiments have a number of notable features. Firstly the sheer size of model in
use is substantially larger than has been previously achieved. The model has state dimension
approximately 4 x 106 and approximately 4 x 104 observations per analysis step. This
is 2 orders of magnitude more than has been achieved with a particle filter in the
geosciences. Secondly, the use of a fully nonlinear data assimilation technique to
initialise a climate model gives us the possibility to find non-Gaussian estimates
for the current state of the climate. In doing so we may find that the same model
may demonstrate multiple likely scenarios for forecasts on a multi-annular/decadal
timescale.
The experiments consider to assimilating artificial sea surface temperatures daily for
several years. We will discuss how an ensemble based method for assimilation in a
coupled system avoids issues faced by variational methods. Practical details of
how the experiments were carried out, specifically the use of the EMPIRE data
assimilation framework, will be discussed. The results from applying the nonlinear data
assimilation method can always be improved through having a better representation of
the model error covariance matrix. We will discuss the representation which we
have used for this matrix, and in particular, how it was generated from the coupled
system. |
|
|
|
|
|
|