A large variety of hydrological responses of catchments exist. This diversity is due to the multiplicity of hydrological processes that govern these responses. Consequently, it is difficult to find a representation that is rich and flexible enough to be adapted to all possible conditions. The hydrological models, which are mathematical representations of hydrological responses built to transform climatic variables (i.e. the rainfall) to hydrological variables (i.e. the river flow), are sometimes limited by their capacity to correctly reproduce these responses. One option to overcome this issue is to combine different models that are complementary in terms of processes representation. These combination methods are called "multimodel approaches". Such methods are based on the mixture of models simulated flows without any interaction between the models. This thesis aims at testing a multimodel approach based on the dynamical coupling of several models by transferring information during the run. The used method, called "supermodel", was set up by climatologists and the present work aims at evaluating its transferability to the context of hydrological modelling. The method relies on the modification of each model's internal states values regarding the values of the other models corresponding internal states (a parallel can be made with a simulated data assimilation technique). The hydrological models used in this thesis are bucket-type models. An implementation of the models into state-space representations was made prior to the set up of the supermodel. This thesis presents two applications of this method using bucket-type rainfall-runoff models: