 |
| Titel |
Joint State and Parameter Estimation in Sequential Data Assimilation with Ensemble Kalman Filter and Meta-Model |
| VerfasserIn |
X. Han, H. J. Hendricks Franssen, X. Li |
| Konferenz |
EGU General Assembly 2012
|
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 14 (2012) |
| Datensatznummer |
250060586
|
|
|
|
|
|
| Zusammenfassung |
| Two methods have become popular to jointly update states and parameters with Ensemble
Kalman Filter (EnKF). In these methods, uncertainties with respect to the model states and
parameters can be treated simultaneously. The two methods are: (1) SODA (Simultaneous
Optimization and Data Assimilation), in which EnKF is combined with an optimization
algorithm to update the model states and parameters separately; (2) Dual state parameter
estimation (Dual), in which the model state vector of EnKF is augmented to include both
model states and parameters in a new vector, which is updated during the assimilation.
Because the SODA method is very CPU-intensive (especially for distributed models), it has
not been applied widely. The Dual method has become a popular approach in recent
applications. However, the disadvantage of the Dual method is that the model results
are very sensitive to the sampling procedure of the high-dimensional parameter
space and that the methodology performs only optimally for Gaussian distributions.
Therefore, there is still potential to improve the joint state and parameter estimation with
EnKF.
Here, a new efficient joint estimation method is proposed. The method consists of two
steps: (1) State update: EnKF is used to assimilate the observations and update the model
states vector; (2) Parameter update: a quadratic response surface regression method is used to
fit a meta-model which describes the relationship between the model parameter
ensemble members and the innovation vector of the EnKF. In addition, new model
parameter ensemble members are generated and used as input to this meta-model
to predict the new innovation vector. Finally the parameters which minimize the
innovation vector are chosen as the updated parameters at the current time step. These
two steps are carried out sequentially according to the availability of observation
data.
The advantage of this new method is that a high dimensional parameter distribution space
can be treated efficiently using the mean-model. This ensures the traversal of potential model
parameter values and quick convergence. The Lorenz63 model was used to evaluate the
new proposed joint estimation method. The results showed that the three model
parameters could converge to the true parameter values in 700 model steps and
the joint state and parameter estimation performed very well with this method. |
|
|
|
|
|
|