Abstract. Ensemble weather forecasts can help to better manage and anticipate the risks of extreme weather events. Nevertheless, weather forecasting is a complex task due to the chaotic behaviour of the atmosphere, which is a major source of uncertainties for sub-seasonal time scale (days to a month). To overcome these uncertainties, a large number of numerical simulations are required. It allows to determine the statistical distribution of the climate variables. In this thesis, we have developed a weather ensemble forecasting tool based on statistical and probabilistic methods to generate weather ensemble forecasts. We used a stochastic weather generator designed to mimic the behaviour of climate variables, based on atmospheric circulation analogs. We have tested this tool to forecast different climate variables such as European precipitation and the Madden-Julian oscillation. We have evaluated the performance of our forecasts using several forecast verification methods. In addition, we compared the performance of our forecast to other forecasts from international weather centers. We start by assessing the capacity of the stochastic weather generator to simulate precipitation in Europe at the local scale (city level). We evaluated the ensemble forecasts of averages over 3 to 60 days ahead. We found good performances in different regions of Europe for up to 10 days. We assessed the role of atmospheric circulation patterns on the forecast scores of meteorological parameters. We also identified the influence of weather regimes on forecast performances. Then, we combined the stochastic weather generator with dynamical model outputs to obtain large ensembles of European precipitation forecasts. We obtain interesting forecast scores for averages up to 35 days ahead at a very local scale. This led to a significant improvement over the forecasts of the European Centre for Medium-Range Weather Forecasts and Météo-France. Finally, we adjusted our model to forecast the Madden Julian Oscillation (MJO). The MJO is responsible for heavy precipitation in densely populated regions such as India. Our model provides a forecast of averages of MJO indices up to 40 days in advance and is competitive with numerical weather predictions. The results of this thesis have been the subject of published scientific papers. Some other work on the predictability of meteorological variables has also been developed.

Mme Sylvie Joussaume – Examinatrice – Directrice de recherche CNRS, LSCE-IPSL
Mme Juliette Blanchet – Rapporteur – Chargée de recherche CNRS, IGE, Grenoble
M. Grégoire Mariéthoz – Rapporteur – Professeur à l’université de Lausanne, Suisse
Mme Lauriane Batté – Examinatrice – Direction de la Climatologie et des Services Climatiques, Météo France
M. Frédéric Vitard – Examinateur – Principal Investigator, ECMWF, Reading, UK
M. Alvaro Corral – Invité – Professeur au Centre de Recerca Matemàtica, Barcelone, Espagne

Sous la direction de :

M. Pascal Yiou – directeur de thèse – Chercheur, CEA-LSCE
Mme Céline Déandreis – co-encadrante de thèse – Ingénieure de recherche, ARIA Technologies