Abstract : Karst aquifers provide water resource for a large part of the Mediterranean population and
water resource becomes a strategic problem during summer when population increases due to
tourism. To help managers to optimize the exploitation of water, this paper studies the ability of a
neural network model to efficiently simulate water levels in the Cèze River, connected to a karst
aquifer, few months ahead during the dry season. The neural model is a recurrent multilayer
perceptron that learns the relations between inputs (mainly rainfall and ETP) and output (water level).
After a training step using 17 years of data, the model is assessed on a never seen year to be
validated. A particular attention is devoted to the dry season (from May to September). The model
achieved good forecast of the maximal observed drawdown. Several architectures were run, each one
related to a specific hypothesis about the behaviour of the hydrosystem or about the strategy of
modelling. Recurrent multilayer perceptrons were used to achieve these tasks and thanks a rigorous
process to variables and model complexity selection, obtained performances were very satisfying.
Assessment of the model was done on one of the drier summer of the database composed of 19
years of daily water-levels. For the five tested architectures, Nash criteria evolve between 0.84 to 0.9
on the whole year, and between 0 to 0.6 on the drier period (June to August). As this modelling
strategy can be fed by rainfall scenario, this kind of models could help managers to optimize
anthropogenic impacts on the river to preserve natural ecosystems. The methodology is generic and
can thus be used with profit by managers on other hydrosystems.
