![Hier klicken, um den Treffer aus der Auswahl zu entfernen](images/unchecked.gif) |
Titel |
Hydrological data assimilation using Extreme Learning Machines |
VerfasserIn |
Marie-Amélie Boucher, John Quilty, Jan Adamowski |
Konferenz |
EGU General Assembly 2017
|
Medientyp |
Artikel
|
Sprache |
en
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 19 (2017) |
Datensatznummer |
250142142
|
Publikation (Nr.) |
EGU/EGU2017-5722.pdf |
|
|
|
Zusammenfassung |
Data assimilation refers to any process that allows for updating state variables in a model to
represent reality more accurately than the initial (open loop) simulation. In hydrology, data
assimilation is often a pre-requisite for forecasting. In practice, many operational agencies
rely on “manual” data assimilation: perturbations are added manually to meteorological
inputs or directly to state variables based on “expert knowledge” until the simulated
streamflow matches the observed streamflow closely. The corrected state variables are then
considered as representative of the “true”, unknown, state of the watershed just before the
forecasting period. However, manual data assimilation raises concerns, mainly
regarding reproducibility and high reliance on “expert knowledge”. For those reasons,
automatic data assimilation methods have been proposed in the literature. Automatic
data assimilation also allows for the assessment and reduction of state variable
uncertainty, which is predominant for short-term streamflow forecasts (e.g. Thiboult et al.
2016).
The goal of this project is to explore the potential of Extreme Learning Machines (ELM,
Zang and Liu 2015) for data assimilation. ELMs are an emerging type of neural network that
does not require iterative optimisation of their weights and biases and therefore are much
faster to calibrate than typical feed-forward backpropagation neural networks. We explore
ELM for updating state variables of the lumped conceptual hydrological model
GR4J. The GR4J model has two state variables: the level of water in the production
and routing reservoirs. Although these two variables are sufficient to describe the
state of a snow-free watershed, they are modelling artifices that are not measurable.
Consequently, their “true” values can only be verified indirectly through a comparison of
simulated and observed streamflow and their values are highly uncertain. GR4J
can also be coupled with the snow model CemaNeige, which adds two other state
variables (snow calorific deficit and the percentage of the watershed covered by
snow).
In order to verify the general applicability of the proposed method, it is applied
to five watersheds in contrasting hydro-climatic contexts (e.g., nordic, semi-arid,
humid, and temperate). We show that the ELMs can be used to successfully update
GR4J’s state variables, in all cases. Considering the ensemble mean as a deterministic
forecast, the Mean Absolute Error between simulated and observed streamflow can
be reduced by up to 30 %. We also show that the success of the method is highly
dependent on adequate selection of input variables for the ELMs. Overall, this
new data assimilation method offers good performance with low computational
costs.
References
Thiboult, A., Anctil, F. and Boucher, M.-A. (2016) Accounting for three sources of
uncertainty in ensemble hydrological forecasting, Hydrology and Earth System Sciences,
20(5), 1809-1825.
Zang, K. and Liu, M. (2015) Outlier-robust extreme learning machine for regression
problems, Neurocomputing, 151, part 3, 1519-1527. |
|
|
|
|
|