Research papersA regional model for extreme rainfall based on weather patterns subsampling

- We model rainfall extremes in the South-East of France.
- Our regional model is based on a seasonal and weather-pattern subsampling.
- Local models with heavy-tailed distributions are not robust.
- An impressive improvement of model robustness is obtained when RFA is applied.
- Rainfall extreme models based on the GPD are discarded for an operational use.

SummaryMany rainfall generators rely on the assumption that statistical properties of rainfall observations can be related to physical processes via weather patterns. The MEWP (Multi-Exponential Weather Pattern) model belongs to this class. In this daily rainfall model, extremes above a threshold are distributed exponentially, for each season and atmospheric circulation pattern. A wide range of applications of this rainfall compound distribution has demonstrated its robustness and reliability. However, recent investigations showed that MEWP tends to underestimate the most extreme rainfall events in specific regions (e.g. the South-East of France).In this paper, we apply different versions of a generalized MEWP model: the MDWP (Multi-Distribution Weather Pattern) model. In the MDWP model, the exponential distribution is replaced by distributions with a heavier tail, such as the Generalized Pareto Distribution (GPD). Unfortunately, local applications of the MDWP model reveal a lack of robustness and overfitting issues. To solve this issue, a regional version of the MDWP model is proposed.Different options of a regionalization approach for excesses are scrutinized (e.g. choice of the scale factor, testing of homogeneous regions based on neighborhoods around each site, choice of the distribution modelling extreme rainfall). We compare the performances of local and regional models on long daily rainfall series covering the southern half of France. These applications show that the local models with heavy-tailed distributions exhibit a lack of robustness. In comparison, an impressive improvement of model robustness is obtained with the regional version, without a loss of reliability.