La prévision des crues des petits bassins versants, avec un modèle pluie-débit, est fortement conditionnée par la connaissance de la pluie. Cette information, estimée par des mesures de pluviographes ou de radar, est entachée de nombreuses incertitudes.Les services français de prévision des crues disposent maintenant d'une version du modèle conceptuel pluie-débit GR3H, adaptée à la prévision opérationnelle. Il utilise une procédure d'optimisation d'un seul paramètre, le niveau initial du réservoir-sol.On a voulu tester le comportement de ce modèle, face à différentes perturbations du signal pluie de base. L'hydrogramme de notre crue de référence a été préalablement simulé avec GR3H.On a montré que le modèle est capable d'absorber d'importantes variations du signal pluie, mais seulement si l'origine et la fin de l'épisode pluvieux sont respectés. Sinon, pour compenser un décalage temporel entre pluie et débit, on a combiné plusieurs modèles GR3H à temps de réaction différents, avec une procédure multimodèles simplifiée. Enfin, pour éviter quelques instabilités, on a testé une variante baptisée "multidélais" qui a permis d'apporter un gain supplémentaire sur la qualité de la prévision., Flood forecasting in small watersheds (some hundreds of km2) has to take into account rainfall. This is why the lumped conceptual rainfall-runoff GR3H model (Cemagref) has been adapted for the French flood forecasting services for operational use. However, the relevance of forecasting is strongly conditioned by the knowledge of real rainfall on the drainage basin. This information, estimated by rain gauge measurements or meteorological radar, contains numerous quantitative and sometimes temporal uncertainties. In this study, we tested the influence of these uncertainties on the behaviour of the GR3H forecasting model.In the GR3H model, the input is the hourly rainfall on the watershed and the output is the hourly flow at the outlet. The production function uses one parameter (A), which represents the higher soil reservoir level. The transfer function uses two parameters: B (the maximal capacity of transfer reservoir) and C (the base time of unit hydrographs HU1 and HU2). In a discontinuous event mode, we have to add an additional parameter S0/A, the initial level of soil reservoir A. For each event, it represents the initial hydrological state of the basin. When used as a forecasting model, A, B and C values are fixed. Thus, to adapt the GR3H model for operational forecasting, we used an optimization process to select the S0/A value. At every moment, this process looks for the S0/A value that makes the calculated discharge equal to the known discharge.To test the impact of rainfall signal perturbations on our forecasting process, we worked on a theoretical flood. Its hydrograph was simulated using the GR3H model from a basic rainfall signal with a constant intensity of 10 mm/h over 12 h. The parameters (A = 400 mm, B = 80 mm, C = 6 h) came from a study of 16 flood events, in a basin of 215 km2, located in the French Pyrenees. To initialize the reservoir levels, base runoff was 1 m3 /s (for reservoir B) and S0/A was fixed at 0.65 (for reservoir A). As an operational scenario, we worked without a precipitation forecast (null future rainfall hypothesis); thus, the forecast time was limited to half the C value, i.e. 3 h, due to the parabolic pattern of the unit hydrograph HU1. To quantify forecast performances, we used the persistence index, which compares the studied model with an inert model (i.e., future is equal to present). We tested successively three kinds of perturbations on rainfall signal:1. the variability (max 50%) of the hourly rainfall intensity, over 50 simulations, preserving the total sum of rainfall;2. the variability (max 50%) of the total sum of rainfall, over 11 cases, preserving a constant intensity (from 5 to 15 mm/h) and3. shifting the beginning time of rainfall, over 7 cases from –3h to +3h.For each kind of perturbation, we considered two forecasting protocols: first a non-operational protocol in which the initial state is known a priori (fixing S0/A); and second as in operational situations, in which the initial state is unknown (optimizing S0/A). We demonstrated that our optimization updating process was rather well adapted to balance the quantitative variability of rainfall. On the other hand, it was not effective to balance an important temporal shift in rainfall. Indeed, in the GR3H model, the temporal parameter (C) is independent from production parameters (A, B and S0/A). To solve this problem we used a multi-model procedure (PMM), i.e., a linear weighting method of results from different forecasting models. The weight of each variable depends on the relevance of the past forecast. We combined three different GR3H models with the same A and B values and different C values (4 h, 6 h and 8 h). This method gave better results but we observed some forecasting instabilities. To solve this problem, we used a multi-time PMM. To improve the 3 h time forecast, we also considered the performances of short-term forecasts (1 h and 2 h). We tested GR3H forecasting over ten French watersheds, using from 12 to 25 events. Results were rather interesting, except when the rainfall signal was not representative of the real spatio-temporal variability (e.g., thunderstorms or basins that were too large). In these cases, semi-distributed models should be useful.A priori, our conclusions were focused on the GR3H model and our updating procedure. However, we propose that they could be similar for other hydrological global models, which use reservoirs and few parameters, offering some inertia and stability to the system. To conclude, when the GR3H was able to model the hydrological behaviour of a small watershed, forecasts were not strongly influenced by quantitative imprecision in the rainfall signal, as long as this imprecision did not greatly affect the beginning and, mainly, the end of the rainy episode.