Machine Learning applied to the Garonne

Description of the Garonne case study:

The hydraulic solver MASCARET ( developed by EDF) is used to build a hydraulic model for a 50-km reach of the Garonne river from Tonneins to La Réole with an observing station at Marmande.
The Garonne River flows from the Pyrenees to the Atlantic Ocean in the area of Bordeaux. It is approximately 647-km long and drains an area of 55,000 km2. The mean slope over the reach is 3.3 and the mean width of the river is 250m. The bank-full discharge is approximately equal to the mean annual discharge (Q = 1000 m3.s-1). Despite the existence of active floodplains, this reach can be modeled accurately by a 1-D hydraulic model.
The hydraulic model for the the Garonne River is built from 83 on-site bathymetry cross sections from which the full 1-D bathymetry is interpolated. Friction is prescribed over three portions. The upstream boundary condition is prescribed with a discharge and the downstream boundary condition is prescribed with a local rating curve established at La Réole. The hydraulic model has been calibrated using channel and floodplain roughness coefficients as free parameters [1].
A MASCARET integration for a typical flood event over 2 to 3 days takes about 30 s CPU time.
This hydraulic test case was distributed by EDF to the hydraulic community as a common framework for simulation, uncertainty quantification and data assimilation investigation.

Description of the Garonne data set:

The present Garonne data set contains hourly water level and discharge discretized over 463 points simulated with MASCARET over the 1996-2004 period using observed discharge at Tonneins and water level at La Réole. The repository contains two folders with inputs at Tonneins and La Réole as well as simulation outputs and a python file to read these data.

This dataset can be downloaded here.

Description of the machine learning objective

The idea is to train a model over the simulation period with observed discharge at Tonneins and water level at La Réole as inputs and simulation from the hydraulic model as a target.
We could then use it as a surrogate model to predict to predict discharge and water level values throughout the entire section, thus replacing the use of the direct solver and drastically reducing the computational cost.

References: [1] Besnard, A., Goutal, N.: Comparaison de modèles 1D à casiers et 2D pour la modélisation hydraulique d'une plaine d'inondation - Cas de la Garonne entre Tonneins et La Réole. La Houille Blanche 3, 42-47 (2011)