![Hier klicken, um den Treffer aus der Auswahl zu entfernen](images/unchecked.gif) |
Titel |
A novel data-driven approach to model error estimation in Data Assimilation |
VerfasserIn |
Sahani Pathiraja, Hamid Moradkhani, Lucy Marshall, Ashish Sharma |
Konferenz |
EGU General Assembly 2016
|
Medientyp |
Artikel
|
Sprache |
en
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 18 (2016) |
Datensatznummer |
250125755
|
Publikation (Nr.) |
EGU/EGU2016-5389.pdf |
|
|
|
Zusammenfassung |
Error characterisation is a fundamental component of Data Assimilation (DA) studies.
Effectively describing model error statistics has been a challenging area, with many
traditional methods requiring some level of subjectivity (for instance in defining the error
covariance structure). Recent advances have focused on removing the need for tuning of error
parameters, although there are still some outstanding issues. Many methods focus only on the
first and second moments, and rely on assuming multivariate Gaussian statistics. We
propose a non-parametric, data-driven framework to estimate the full distributional
form of model error, ie. the transition density p(xt|xt−1). All sources of uncertainty
associated with the model simulations are considered, without needing to assign error
characteristics/devise stochastic perturbations for individual components of model
uncertainty (eg. input, parameter and structural). A training period is used to derive
the error distribution of observed variables, conditioned on (potentially hidden)
states. Errors in hidden states are estimated from the conditional distribution of
observed variables using non-linear optimization. The framework is discussed in
detail, and an application to a hydrologic case study with hidden states for one-day
ahead streamflow prediction is presented. Results demonstrate improved predictions
and more realistic uncertainty bounds compared to a standard tuning approach. |
|
|
|
|
|