Résultats du projet Extraflo sur la comparaison des méthodes d'estimation des pluies extrêmes en France

This paper focuses on a comparison of 3 families of extreme rainfall frequency analysis methods: 1) purely local approaches which consist in fitting a probability distribution function (pdf) to a sample of observed daily rainfall, 2) local-regional approaches where the sample of annual maxima daily rainfall at the target site is enlarged by rainfall observations in a surrounding homogeneous neighbourhood and 3) purely regional approaches for estimating the extreme rainfall pdf at an ungaged site. For 2) and 3) the extreme rainfall pdf can be estimated either by frequency analysis based on the extreme value theory or from the simulations of the French stochastical rainfall model named SHYREG. The comparison relies on a dataset of 1568 daily rainfall gages belonging to the French Weather Service (Meteo-France) or the French Electricity Company (EDF) in the southern half of France. A comparison methodology and several comparison criteria have been defined to evaluate the bias and variances of the quantiles and the confidence intervals. The main results are: 1) At the local scale, the General Pareto Distribution fitted on a peak over threshold sample performs better that the Generalized Extreme Value pdf fitted on max annual sample, 2) the exponential pdf is clearly unsuitable in Mediterranean area, 3 the local approaches perform poorly compared to the local-regional and purely regional frequency analyses except for the MEWP weather pattern frequency analysis which is comparable to the local regional frequency analysis for non-Mediterranean climate and when the calibration sample size exceeds 10 years, 4) the local regional frequency analysis and the SHYREG method performances are comparable, and 5) there is only a weak decrease in performances when using a purely regional frequency analysis instead of a local-regional frequency analysis