A. Randrianasolo (1), M.H. Ramos (1), G. Thirel (2), V. Andréassian (1), E. Martin (2)

1. 1 Objective: nnumber of forecast members Nnumber of days used to compute the score ximean of the ensemble forecasts for the day i xk,i value of the member k for the day i Nnumber of days used to compute the score xj= 0 (the event occurs) xj= 1(the event does not occur) pj probability of the event to occur 2 Data and methods: 4 Results Fig. 4: Ratio-RMSE values (Leadtime = Day 2) Fig. 1: Location of the catchments 5 Conclusions Impact of the use of two different hydrological models on scores of hydrological ensemble forecasts A. Randrianasolo (1), M.H. Ramos (1), G. Thirel (2), V. Andréassian (1), E. Martin (2) • Hydrology Research Group, Cemagref HBAN, Antony, France (contact: maria-helena.ramos@cemagref.fr) • CNRM-GAME, Météo-France, CNRS, GMME/MOSAYC, Toulouse, France (contact: guillaume.thirel@meteo.fr ) to assess the quality of ensemble streamflow forecasts issued by two different modelling conceptualizations of catchment response, both driven by the same weather ensemble prediction system • 211 catchments in France (170 to 9390 km2) (Fig. 1) • Weather forecasts from the PEARP ensemble prediction system of Météo-France (March 2005-July 2006): • 11 perturbed members for a forecast range of 60 h (skill scores computed for the first two days of forecast range) • Time series of observed data: daily discharge, precipitation, temperature • Two hydrological models: • the coupled physically-based hydro-meteorological model SAFRAN-ISBA-MODCOU developed at Météo-France, based on a distributed catchment model (Fig. 2a) Fig. 3: POD, FAR, BIAS (Leadtime = Day 1 and Qref2 = Q90) GR with updating SIM GR without updating 2) the lumped soil-moisture-accounting type rainfall-runoff model (GR3P) developed at Cemagref (Fig.2b) Fig. 2a: SIM model Fig. 2b: GR3P model catchment area (km2) GR with updating SIM GR without updating www.cemagref.fr/webgr Fig. 5: Spread Fig. 6: BSS Q90 Q50 3 Skill scores Day 1 • Two critical thresholds for observed events: Qref1 = 50th percentile (Q50) Qref2 = 90th percentile (Q90) Contingency table • Threshold for forecasted events: if p = 50% of the members are greater than Qref, the event is considered as a « forecasted event » Day 2 Standard deviation (or spread) Brier Skill Score Root Mean Square Error oi observed data for the day i mi mean of the ensemble forecasts for the day i N number of days used to compute the score Ratio-RMSE: RMSE / Mean of observed streamflows Ratio-σ: Standard Deviation / Mean of forecasted streamflow BSS: the reference used is the climatology • PEARP-based ensemble streamflow forecasts predicted well discharges over the studied catchments • Better scores are obtained from the GR3P model with updating, while SIM results are closer to the results from GR3P model without updating (for data assimilation in SIM model, see Thirel et al., 2009) References: 1. Thirel, G., Rousset-Regimbeau, F., Martin, E., Habets, F. (2008) On the impact of short-range meteorological forecasts for ensemble streamflow predictions. J. Hydrometeorology (9), 1301-1317. 2. Tangara, M. (2005) Nouvelle méthode de prévision de crue utilisant un modèle pluie-débit global. PhD Thesis EPHE-Cemagref, Paris, 374 p. 3. Thirel, G., E. Martin, J. F. Mahfouf, S. Massart, S. Ricci, and F. Habets (2009) A streamflow assimilation system for ensemble streamflow forecast over France. Abstract EGU2009-6890. 4. Randrianasolo, A. (2009) Evaluation de la qualité des prévisions pour l'alerte aux crues. MSc Thesis ENGREF, Cemagref (ongoing). HEPEX09 Workshop – Toulouse, 15-18 June 2009