In this work, we implemented the GHER hydrodynamic model in the Gulf of
Lions (resolution 1/100°). This model is nested interactively in another model
covering the North-Western basin of the Mediterranean Sea (resolution 1/20°),
itself nested in a model covering the whole basin (1/4°). A data assimilation
filter, called the SEEK filter, is used to test in which of those grids observations taken in the Gulf of Lions are best assimilated. Therefore, twin experiments are used: a reference run is considered as the truth, and another run, starting from different initial conditions, assimilates pseudo-observations coming from
the reference run. It appeared that, in order to best constrain the coastal
model, available data should be assimilated in that model. The most efficient setup, however, is to group all the state vectors from the 3 grids into a single vector, and hence coherently modify the 3 domains at once during assimilation cycles.
Operational forecasting with nested models often only uses so-called passive
nesting: no data feedback happens from the regional models to the global model.
We propose a new idea: to use data assimilation as a substitute for the feedback.
Using again twin experiments, we show that when assimilating outputs
from the regional model in the global model, this has benecial impacts for the
subsequent forecasts in the regional model.
The data assimilation method used in those experiments corrects errors in the
models using only some privileged directions in the state space. Furthermore, these directions are selected from a previous model run. This is a weakness of the method when real observations are available. We tried to build new directions of the state space using an ensemble run, this time covering only the Mediterranean basin (without grid nesting). This led to a quantitative characterization of the forecast errors we might expect when various parameters and external forcings are affected by uncertainties.
Finally, using these new directions, we tried to build a statistical model supposed to simulate the hydrodynamical model using only a fraction of the computer resources needed by the latter. To achieve this goal, we tried out artifficial neural networks, nearest-neighbor and regression trees. This study constitutes only the first step toward an innovative statistical model, as in its present form, only a few degrees of freedom are considered and the primitive equation model is still required to build the AL method. We tried forecasting at 2 different time
horizons: one day and one week.