نبذة مختصرة : Karst is one of the most widespread aquifer formations in the world. Their exploitation provides fresh water to practically 25% of the global population. The high level of structure heterogeneity in these aquifers however makes them complex and their behavior is difficult to study, simulate and forecast. Artificial neural networks are machine learning models widely used in surface hydrology since the 90's thanks to their properties of parsimony and universal approximation. In this thesis, artificial neural networks are used to study karst aquifer behavior. Application is done on the Lez. This aquifer situated near Montpellier conurbation (400 000 inhabitants) provides fresh water for a large part of this population. First, a "classical" black box neural network is applied to simulate and forecast Lez spring discharge. A rainfall input selection method is proposed, using cross correlation analysis and cross validation method at the same time. Results show neural model efficiency in order to simulate and forecast the spring discharge of a complex karstic aquifer. The model was tested using two hydrologic cycles including the two most intense floods of the database. Hydrographs shows that neural model was able to forecast correctly the maximum flood discharge of these intense floods when they are higher than all discharges of the learning database. Forecasting is satisfactory until a one-day horizon. In a second time, extraction of the knowledge included in the black box is proposed. In order to constrain the model to provide physically plausible solution, a priori knowledge about aquifer geology is included into the network architecture. KnoX (Knowledge eXtraction) method proposed in this study aims at extract geological zone contributions to the Lez spring and corresponding response times. The KnoX methodology was applied to a fictitious hydrosystem built using a model with controlled parameters, in particular contributions of subbasin to the outlet and lag time of each subbasin. This application permitted to ...
No Comments.