|
Titel |
Performance comparison of different EnKF variants is far from trivial |
VerfasserIn |
Johannes Keller, Harrie-Jan Hendricks-Franssen, Gabriele Marquart |
Konferenz |
EGU General Assembly 2016
|
Medientyp |
Artikel
|
Sprache |
en
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 18 (2016) |
Datensatznummer |
250122380
|
Publikation (Nr.) |
EGU/EGU2016-1402.pdf |
|
|
|
Zusammenfassung |
%-----------Document------------------------
\begin{document}
The estimation of permeability values alongside their uncertainty is
vital in the assessment of groundwater flow and heat transport. The
Ensemble Kalman Filter (EnKF), an efficient tool for sequential data
assimilation, has proved to be useful in estimation of parameters that
display complex spatial patterns. Several variants of the EnKF have
been proposed to improve the assimilation for small ensemble sizes,
highly nonlinear systems or with non-Gaussian probability
distributions. In this study, we evaluate the performance of different
EnKF-methods in runs with $50, 100, 250, 500, 1000$ and $2000$
ensemble members. These experiments are carried out for eight
different variants of EnKF and repeated $1000$ times.
The computation of a tracer test in a 2D heterogeneous permeability
model provided the measurement values for the EnKF. The considered
EnKF methods include the standard EnKF, a damped EnKF, EnKF including
localization, a Normal Score EnKF, a Hybrid EnKF scheme, a Dual EnKF,
iterative EnKF and square root EnKF. For each simulation the Root Mean
Square Error was recorded yielding a RMSE distribution for the $1000$
synthetic experiments.
The main outcome from the simulation experiments was that the
comparison of different EnKF-variants needs many repetitions in the
form of different synthetic tests, even if the ensemble size is large,
in order to reach a sound conclusion. In addition, it was found that
for a small ensemble size of $50$ members most EnKF-variants
outperformed classical EnKF. On the other hand, for large ensembles
sizes differences between methods were smaller and classical EnKF was
hardly outperformed. EnKF with dampening and iterative EnKF gave the
best results.
\end{document} |
|
|
|
|
|