What contributions of a semi-distributed multi-model approach for streamflow forecasting ? ; Quels apports d'une approche multi-modèle semi-distribuée pour la prévision des débits ?

Streamflow forecasting is needed to meet various objectives, from the safety of populations in the case of floods, to the management of water resources for multiple uses, particularly important during low flow periods. Hydroelectric production is also closely linked to streamflow, as is the case on the Rhône catchment, where the Compagnie Nationale du Rhône (CNR) manages hydroelectric production units along the main course of the river. To optimize its production, CNR relies in part on hydrological models, which transform meteorological information into streamflow estimates. Hydrological models have their own specificities, in terms of the way they represent the catchment (i.e. their mathematical structure) or the way they are applied, linked to the objectives or hydro-climatic contexts for which they were designed. It leads to a very wide range of hydrological models, representing various rainfall-runoff relationships and operating at different spatial scales, often making it difficult for users to choose the structure and spatial discretization best suited to their objectives. This also reflects the numerous uncertainties affecting the modelling chain, the quantification of which is of major importance for operational management and decision support. In this context, the aim of this PhD is to evaluate the contribution of a semi-distributed multi-model approach for streamflow forecasting. To this end, a large sample of 643 French catchments, with hydro-climatic data at hourly time steps, was compiled, and 14 hydrological model structures were used according to different calibration strategies and for different spatial configurations. First, we set up the hydrological modelling framework to meet our objectives. After that, we tested the semi-distributed multi-model approach in a simulation context, and then for streamflow forecasting purposes, focusing on the case of Rhône tributaries. The results show that the strategy of a single model for a large number of catchments has limitations that can be partly ...