This PhD work proposes new concepts and tools for stochastic weather simulation activities targeting the specific needs of hydrology. We used, as a demonstration, a climatically contrasted area in the South-East of France, Cévennes-Vivarais, which is highly attractive to hydrological hazards and climate change.Our perspective is that physical features (soil moisture, discharge) relevant to everyday concerns (water resources assessment and/or hydrological hazard) are directly linked to the atmospheric variability at the basins scale, meaning firstly that relevant time and space scales ranges must be respected in the rainfall simulation technique. Since hydrological purposes are the target, other near-surface variates must be also considered. They may exhibit a less striking variability, but it does exist. To build the multi-variable modeling, co-variability with rainfall is first considered.The first step of the PhD work is dedicated to take into account the heterogeneity of the precipitation within the rainfall simulator SAMPO [Leblois and Creutin, 2013]. We cluster time steps into rainfall types organized in time. Two approaches are tested for simulation: a semi-Markov simulation and a resampling of the historical rainfall types sequence. Thanks to clustering, all kind of rainfall is served by some specific rainfall type. In a larger area, where the assumption of climatic homogeneity is not considered valid, a coordination must be introduced between the rainfall type sequences over delineated sub-areas, forming rainy patterns at the larger scale.We first investigated a coordination of Markov models, enforcing observed lengths-of-stay by a greedy algorithm. This approach respects long duration aggregates and inter-annual variability, but the high values of rainfall are too low. As contrast, the joint resampling of historically observed sequences is easier to implement and gives a satisfactory behavior for short term variability. However it lacks inter-annual variability.Both approaches suffer from the strict delineation of homogeneous zones and homogeneous rainfall types.For these reasons, a completely different approach is also considered, where the areal rainfall totals are jointly modeled using a spatio-temporal copula approach, then disaggregated to the user grid using a non-deterministic, geostatistically-based conditional simulation technique. In the copula approach, the well-known problem of rainfall having atom at zero is handled in replacing historical rainfall by an appropriated atmospheric based rainfall index having a continuous distribution. Simulated values of this index can be turned to rainfall by quantile-quantile mapping.Finally, the copula technique is used to link other meteorological variables (i.e. temperature, solar radiation, humidity, wind speed) to rainfall. Since the multivariate simulation aims to be driven by the rainfall simulation, the copula needs to be run in conditional mode. The achieved toolbox has already been used in scientific explorations, it is now available for testing in real-size application. As a data-driven approach, it is also adaptable to other climatic conditions. The presence of atmospheric precursors a large scale values in some key steps may enable the simulation tools to be converted into a climate simulation disaggregation.