Abstract : Precipitations, in particular the rain, constitute a natural phenomenon which has a very strong socio-economical impact, especially when they are torrential feature. To take into account this aspect, the hydrological systems of alert and forecast need more detailed spacetime information and reliable forecast for precipitations in the very short term. This has a particular importance in emergencies (flash flood, urban drainage network management, dam management, etc.). The fields of clouds and precipitations remain the fields most difficult to simulate for the current weather forecasting models. Indeed, the space-time scales of these models remain largely higher than those which are relevant for precipitations: the mechanisms of precipitations are mostly parameterised and the rains are estimated only on relatively large scales. Furthermore, the long spin-up time of these models impede to deliver short-term forecasts. Various statistical methods of processing of satellite and radar images have been developed to make up this deficit of forecast. These methods take into account a great number of information on a small scale, but they do not have a physical base, in particular they do not take into account the strongly nonlinear dynamics of the stormy cells. An alternative which allow a priori to exceed, using the multifractal methods, the limits of the preceding methods was recently considered. It is based on the cascade models and takes into account the hierarchy of the structures as well as their nonlinear interactions over a wide range of space-time scales, the anisotropy between space and time, and causality. Basically, the cascade processes develop gradients of more and more great water contents on more and more small fractions of physical space. This type of models empirically has the advantage of to have a very limited number of parameters which have a strong physical significance and can be evaluated either theoretically or empirically. In this thesis we present the implementation of a procedure corresponding to this alternative and its application to the event from September 8-9, 2002 in Nimes, using radar data provided by the Direction of Climatology of Meteo-France, to determinate their multifractal characteristics. We present also the implementation of a procedure for simulation and forecast of multifractal rain fields and the study of the law or predictability loss.