Résumé
In Mediterranean regions, the difficulty of measuring and forecasting rainfall intensity, as well as the difficulty of identifying flood-generating processes, often lead to the use of statistical models, like neural networks. However, without the coupling with meteorological forecasts, current hydrological models are most often limited to a lead time equivalent to the response time of the basin, i.e. a few hours for small basins. The challenge is to increase this lead time, which is often too short for crisis management. As support for demonstration, a flood forecasting model for the Gardon de Mialet basin (Southern France) has been developed. The neural network model implemented (a Multilayer Perceptron) has been fed with the AROME high resolution model to produce forecasts with lead times of up to 24 hours, compared with 2 to 3 hours previously. The use of raw data predicted by AROME for the 47 rainfall events in the database produced flow results which were used as a basis for improvement. We therefore began by identifying the sources of forecast error (meteorological or hydrological models, or both models together), and then qualifying and quantifying meteorological forecast errors. At the end of this process, an approach is proposed for making an optimal use of weather forecasts from AROME, taking into account errors in synchronization, location and intensity of the rainfall.