Flood risk assessment on small catchment with ECMWF ensemble forecasts in operational conditions Debarre C. Belleudy A. Zuber F. - MTE/SCHAPI Goals - Better anticipate flood risk - Internal organisation - Medium term flood risk map - Goals evoluted during study from discharge forecasting to risk assessment Empirical uncertainty processor Rank histogram 108-120h 2500 2000 1500 1000 500 0 0 1000 2000 3000 4000 5000 6000 GRP GRP+OTAMIN - OTAMIN : [Bourgin, F. (2014)] - Improves reliability - Losing of temporal coherency - No improvement of event detection Bourgin, F. (2014). Comment quantifier l’incertitude prédictive en modélisation hydrologique?: Travail exploratoire sur un grand échantillon de bassins versants (Doctoral dissertation, AgroParisTech) Method Ensemble rainfall forecast ECMWF 50 members Catchement areal rainfall calculation GRP Rainfall runoff model 50 discharge members No post processing Empirical uncertainty processor Discrimination tree Logistic regression Post-processing Reliability Rank histogram POD/FAR Evaluation Raw ensemble and statistical post-processing - Predict risk of flood event - Discharge>Threshold - Predictors = 50 daily maximum discharge from GRP ensemble sorted by values - Raw ensemble = P(Q>Threshold) - Logistic regression & CART : predictor selection & cross-validation Study period 2017-2020 GRP : continuous - lumped - storage type rainfall runoff model from INRAE One catchment : le Grand Morin à Pommeuse (770 km²) response time ~16h Conclusion - Challenge of rare event prediction - Can detect ~75% of the events with ~75% of false alarm - No improvement with statistical post processing on test sample Results To be continued - More catchments : mediteranean area... - Spatial aggregation - Longer study period