10.4122/1.1000000492
Honnorat, Marc
Marc
Honnorat
marc.honnorat@imag.fr
Le Dimet, François-Xavier
François-Xavier
Le Dimet
francois-xavier.le-dimet@imag.fr
Monnier, Jérôme
Jérôme
Monnier
jerome.monnier@imag.fr
Honnorat, Marc
Marc
Honnorat
marc.honnorat@imag.fr
Variational data assimilation of lagrangian data for fluvial hydraulics simulations.
XVI International Conference on Computational Methods in Water Resources
2006
2006
A major difficulty in the simulation of river hydraulics flows is bound to model
parameters definition. Solutions of the shallow water equations are determined by
initial conditions, boundary conditions, bed elevation, physical and numerical
parameters. Data assimilation methods make it possible to combine optimally physical
information from the model and observation data of the physical system to identify
the value of model inputs that correspond to a numerical simulation which is
consistent with reality.
Variational data assimilation consists in finding the control variables that minimize
a cost function measuring the discrepancy between the model state variable and
observation data of the physical system. An efficient minimization using a
Quasi-Newton algorithm requires the computation of the gradient of the cost function.
The latter is easily computed from the adjoint state which is solution of an adjoint
model.
However, in river hydraulics, observation data are available only in very small
quantities. Local water level measurements are usually very sparse in space and
velocity measurements are even rarer, especially in case of extreme events such as
floods. Consequently, in practice these eulerian observations are usually not
sufficient to take advantage of data assimilation.
We present a method to use lagrangian data from remote sensing observation in the
assimilation process, in addition to classical eulerian observations of the flow. The
trajectory of particles advected by the flow can bring information on the surface
velocity thanks to an appropriate transport model. Numerical twin experiments
demonstrate that this additional information makes it possible to improve the
identification of model parameters. In order to deal with real data, an observation
operator based on a multi-scale filtering scheme is proposed.