Abstract
The Mediterranean region is frequently subjected to intense rainfalls leading to flash floods. This phenomenon can cause casualties and huge material damages. Facing to this phenomenon, hydrologic forecasting is a major tool used by Service Central d’Hydrométéorologie et d’Appui à la Prévision des Inondations to produce flood warning.During past decades, artificial neural networks showed their efficiency for flash flood forecasting on different type of watershed. The present thesis aims thus to contribute to the development of a generic methodology to design artificial neural networks, that is tested on Gardon d’Anduze and Lez at Lavalette watersheds, both displaying non-linear hydrodynamic behavior. To reduce uncertainties on forecasts, ensemble models, based on the median of forecasts calculated at each time step for an adequate number of models varying only by their initialization, have been proposed. In addition, in order to improve forecasting performances on Gardon d’Anduze, with artificial neural networks, we tried to introduce knowledge about the state of the watersheds before and during the flood. Several variables have thus been tested each one its turn, to select the one given the best performances. On the Lez karst system, that has a strongly heterogeneous structure, the KnoX method have been applicated in order to estimate the contribution to outflow from four geographical zones displaying hydrologic and hydrogeologic behavior considered as homogeneous. Thus, the most contributive zones to the discharge zones have been identified. This will help the investigation of representing humidity variables in these zones.The performances of models underlined that the general methodology of rainfall-runoff model conception could be applied on both basins, even though their hydrological and hydrogeological behavior are very different.The contribution of each zone, estimated from the KnoX methodology, improved comprehension of Lez karst system during flash floods. Selection of relevant variables representing the state of the Lez hydrosystem will be possible thanks to this new knowledge. Performances of models developed in this study underlined the difficulty to find satisfactory models, and showed the interest of the generic methodology used to design neural network adapted to the two targeted basins.