« Improvement of ensemble streamflow predictions over France of the SAFRAN-ISBA-MODCOU model » Guillaume Thirel (CNRM-GAME/GMME/MOSAYC) PhD Director :

« Improvement of ensemble streamflow predictions over France of the

SAFRAN-ISBA-MODCOU model »

Guillaume Thirel (CNRM-GAME/GMME/MOSAYC)

PhD Director : Éric Martin

Jury :

President : Serge Chauzy (LA)

Reviewer : Vincent Fortin (Environment Canada)

Reviewer : Vazken Andréassian (CEMAGREF)

Examiner : Olivier Thual (CERFACS)

Examiner : Pierre Ribstein (UMR Sisyphe)

Context

Floods = major environmental hazard

Damages on infrastructures, huge costs, human beings losses

Flood of the Garonne river at Toulouse in 1875

⇒ Need to better anticipate these events

Organisms (SCHAPI, Services de Prévision des Crues)

Hydrological models

Meteorological forecasts

Context

Ensemble meteorological

forecasts

Post-treatment

Surface observations (snow, discharges, …)

Data assimilation

Hydrological model(s)

Forecasted discharges

Discharges calibration

(from Schaake et al., 2007)

Initial states

Meteorological forecasts

ISBA

Physiographic data pour the soil and the vegetation

+

MODCOU

QrQi

E

H

G

Aquifer

Daily discharges

Surface scheme

Snow

SAFRANObservations + NWP outputs

Precipitation, température, humidity, wind, radiations

Hydrological model

Meteorological analysis

The SIM hydro-meteorological model

Distributed model

Coherent simulation of water and energy fluxes on :• Atmosphere• Surface/vegetation/surface soil• Surface and sub-surface hydrology

Grid mesh : 8x8 km

→ Co-operation Mines Paris Tech /SISYPHECo-operation Mines Paris Tech /SISYPHE

Validations and valorisation of SIM

Validation of the simulations by meteorological and hydrological variables• Snow• River discharges and aquifer levels

Main applications :• Follow-up of soil hydric states, effective rainfall, snow conditions• Impact of climate change• Flood prediction (soil wetness, discharges)

Soil Water Index on 16/11/2009 Direction de la climatologie

Application of SIM to ensemble streamflow predictions

Since 2004, everyday : ensemble discharge forecasts based on SIM (Fabienne Rousset-Regimbeau PhD, 2007). Based on the ECMWF EPS (precipitation+temperature) On the whole France, mid-term range (10 days)

Statistical analysis of precipitations and discharges Article Rousset, ECMWF newsletter spring 2007 Disaggregation of precipitations on a simple, but efficient way Discharges compared to a reference SIM simulation

Study case on a few recent floods

Scheme of the ensemble discharge forecast system based on SIM

Observations + Meteorological

models

SIM ANALYSIS (daily)

SAFRAN10-year

Climatology Wind, Rad.,

Humidity

SOILAQUIFERS

RIVERS

ECMWF/PEARP EPSs 51/11 members, 10/2.5 days forecasts

ENSEMBLE PREDICTIONS

Spatial DISAGGREGATION

T + Precipitations

ISBA MODCOU

ENSEMBLE FORECASTS

SOIL AQUIFERS

RIVERS

ISBA MODCOU

SOIL AQUIFERS

RIVERS

The Seine at Paris, March 2001 flood (decade flood)

Q90

Q50Q10

PhD Fabienne Rousset-Regimbeau

Objectives

To improve the ensemble discharge forecast system

To explore the contribution of 2 EPSs To test an improvement of the model Qualify the chain in comparison with discharge observations

How : By comparing the impact of 2 EPSs on 2-day ensemble discharge

forecasts By improving the system with a past discharge assimilation system

Plan

I Study : comparison of the impact of 2 EPSs in the SIM-based ensemble discharge forecast system

II Past discharges assimilation– 1) Justification

– 2) Choice of the method

– 3) Validation of the data assimilation system

III Impact of the past discharges assimilation system on the ensemble discharges forecasts

IV General conclusions and perspectives

Plan







The 2 used EPSs

ECMWF EPS 51 members 10-day forecasts Singular vectors,

– Optimisation in 48H Resolution in our operational

database : 1.5º

PEARP EPS 11 members 2.5-day forecasts Singular vectors

– Optimisation in 12H Resolution in our operational

database : 0.25°

-> Objective : mid-term range -> Objectif : short-term range

The comparison is done on the first 48H common to both systems

Precipitations disaggregation

Interpolation on the SAFRAN zones according to distance, then :

ECMWF EPS : altitudinal gradient PEARP EPS : correction of the mean bias point by point

SAFRAN ECMWF EPS (Day 1)

PEARP EPS (Day 1)

Precipitation amounts 11 March 2005 / 30 September 2006

All the statistical scores were better for the PEARP EPS

Conclusions on the comparison

The ensemble discharges forecasts based on the PEARP EPS showed an improvement on small basins and for floods– Results confirmed by a set of statistical scores (RPSS, reliability

diagram, False Alarm Rate and Probability of Detection, seasonal study)– Low spread, reference used = SIM simulation– Interest for flood forecasting at a short-term range in France (SCHAPI)

Details of the study in On the impacts of short-range meteorological forecasts for ensemble streamflow predictions, G. Thirel, F. Rousset-Regimbeau, E. Martin, F. Habets, Journal of Hydrometeorology, 2008.

Plan







Justification

Choice of the observations : – Snow : concerns only a limited part of the territory and discharges are

influenced

– Aquifer layers : many data but only few aquifers simulated into SIM

– River discharges : many data over all of France available daily

Choice of the variable to modify : – River water content : efficient for the short-term range, less for the mid-

term range

– Soil water content : concerns the whole territory, impact until the mid-term

Strategy

186 stations assimilated over France– Low human influence

– Good quality of observations (Banque Hydro)

– Good quality of SIM simulations

Principle : to use observed discharges to improve the discharges simulations, by adjusting the ISBA soil moisture

Plan







The BLUE (Best Linear Unbiased Estimator)

Analysed state

Background state Innovation

vector

Observed discharges

Choice of the BLUE because :

Low dimensions of the problem

Possibility to compute the solution in its matricial form

Hypothesis : unbiased errors and linear model

Determination of the K matrix components

To estimate the observations (R) and background covariance errors (B) matrices and calibrate these two matrices between them

To define the state variable : the ISBA soil moisture, but which one?

To estimate the Jacobian matrix H

x

yH

Discharges

Soil moisture

ISBA physics

Runoff : Dunne Subgrid depending on the fraction

of the mesh saturated

Drainage : gravitational subgrid

Improvement of the hydrological transfers in the soil (Decharme et al., 2006; Quintana Seguí et al., 2009)

Discharges : coming from ISBA runoff and drainage

State variable

3 possible choices :

Soil water content : w2+w3(runoff + drainage)

Root zone water content : w2(runoff)

2 soil layers water contents separately : (w2,w3)(runoff and drainage)

Spatial aggregation (sum of the soil water contents over each sub-basin)

Sensitivity of the Jacobian

Perturbation of 1% : the Jacobian varies according to the sign

Perturbation of 0.1% : low modification according to the sign

Thus, we chose to apply a perturbation of +0.1%

-> respect of the linearity

Clear temporal evolution : the Jacobian will be re-calculated for each assimilation

Jacobiennes 0.1%

0

1

2

3

4

5

6

7

sept

-05

oct-0

5

nov-

05

janv-

06

mars

-06

avr-0

6

mai-0

6

juin-

06

juil-0

6

Jaco

bie

nn

e

H J1 +0,1%

H J1 -0.1%

Filling of the Jacobian matrix

3 gauging stations y1, y2 and y3.

x1, x2 and x3 soil water contents summed on the sub-basins

0

0 0

0

sub-basins

stations

Jacobian H :

discharges

Soil moisture

Finite differences

j

iij x

yH

Principle of the assimilation system

Implementation of the assimilation system

PALM coupler (CERFACS) : dynamical coupler dynamique of parallel calculation codes, many applications (data assimilation, coupling)

Friendly interface, modular software Intuitive gestion of data exchanges, buffer storage -> few

modifications of the ISBA and MODCOU codes Simple cluster coupling Use of the Météo-France super-computer

Plan







Assimilation of real observations

6 experiments : 3 state variables * 2 physics of the model Daily assimilation, daily observations

Period : 10 March 2005 / 30 September 2006

186 assimilated stations

The Doubs river at Besançon

-> experiment (modification of the layers 2 and 3 soil moistures + improved physics)

2EI

combines the best Nash and RMSE scores, as well as the lowest increments

(soil moisture + improved physics) will be kept

Scores for 148 assimilated stations

Scores for 49 independent stations

2EI

2EI

Conclusion on the discharges assimilation system

Observed discharges assimilated for the first time in SIM– Positive impact of the use of PALM : CPU time save (parallel computation on

the Météo-France super-computer), modularity

Validation of the assimilation system– System validated on SIM-analysis– Assimilation of real observations : several configurations tested, significative

improvement of the scores, low increments

Article in preparation

For initializing the ensemble discharges forecasts, we will keep :

State variable : mean of the soil moisture into the 2 ISBA layers

The assimilated states (assimilation + improved physics) daily2EI

Perspectives of improvement of the assimilation system

Improvement of the background and observations errors

Reduction of the number of sub-basins in a sub-basin– Less simulations needed for computing H

Tests of other assimilation methods– External loop? (i.e. re-calculating the Jacobian around the analysed state

until it converges) -> tests showed low improvements

– EnKF?

Plan







Conditions of the study

Studied period : 11 March 2005 – 30 September 2006 Scores on 148 assimilated stations Use of the 10-day ECMWF EPS

3 systems of ensemble discharges forecasts were compared : – The real-time system

– A re-forecast initialized by the initial states (modification of the soil moisture of both layers, without the improved physics)

– A re-forecast initialized by the initial states (modification of the soil moisture of both layers, with the improved physics)

2EI

1EI

Some statistical scores

Spread

Ratio-dispersion

0

0,1

0,2

0,3

0,4

0,5

1 2 3 4 5 6 7 8 9 10

Jours

Ratio-dispersionsans assimilation

Ratio-dispersion EI1

Ratio-dispersion EI2

RMSE

Scores computed in comparison with observations

Ratio-RMSE

0

0,2

0,4

0,6

0,8

1

1,2

1,4

1 2 3 4 5 6 7 8 9 10

Jours

Ratio-RMSE sansassimilation

Ratio-RMSE EI1

Ratio-RMSE EI2

Brier Skill Score day 1

Perfect model

Clima-tology

-1

0

1

Q99 Q98 Q95 Q90 Q80 Q70 Q60 Q50 Q40 Q30 Q20 Q10 Q5 Q2 Q1

Quantiles

BSS J1 sans assimilation

BSS J1 EI1

BSS J1 EI2

Brier Skill Score day 10

Perfect model

Clima-tology

-1

0

1

Q99 Q98 Q95 Q90 Q80 Q70 Q60 Q50 Q40 Q30 Q20 Q10 Q5 Q2 Q1

Quantiles

BSS J10 sansassimilationBSS J10 EI1

BSS J10 EI2

Conclusion on the impact of the assimilation

Intrinsic characteristics of the ensemble discharges few modified (spread)

Significative impact of the assimilation for the first days, less important then

Then, the physics improvement improves the forecast quality

Use of the forecasts by the forecasts eased (False Alarm Rates, POD)

Article in preparation

SIM-PEARP less impacted than SIM-ECMWF, scores very close

Plan







General conclusions and perspectives

Two ensemble discharges forecasts systems based on SIM– Impact of the PEARP EPS at a short-range, on small basins and for floods

A past discharge assimilation system implemented in SIM– Validation : significative impact on SIM-analysis

– Low non-linearities

Impact on the ensemble discharges– Strong impact of the assimilation system at a short-range, then low impact

– But the improvement of the physics allows better forecasts at a mid-term range

Perspectives

Implementation of the assimilation system for initializing the operational SIM-ECMWF chain in real-time

Adding aquifer layers in SIM, and then assimilation of aquifer levels (PhD UMR SISYPHE Alexandra Stouls)

Improvement of the meteorological uncertainty taking into account (EPS disaggregation)

Taking into account of uncertainties linked to hydrology : into the initialization and via a stochastic physics or a multi-model forecast

Seasonal forecasts with SIM (PhD CNRM Stéphanie Singla)

My work here on EFAS

Use of satellital snow data for improving the proxy– Particule filter and EnKF

Study of its impact on the EFAS forecasts– Probabilistic statistical scores

2nd step : to see how to use other sources of rainfall data in order to improve the proxy

Thank you for your attention!

Visualisation des sorties en temps réel

Site intramet : http://intra.cnrm.meteo.fr/pedeb/

Sélection d’environ 100 stations

- prévision de débits

- tableau d’alerte

=> Visualisation du risque + de la persistance (ou non) de la prévision

Probabilité de dépassement du seuil d’alerte

http://intra.cnrm.meteo.fr/pedeb/

BSS hauts débits (Q90)

Jour 1 Jour 2

CEPMMT : 49 stations

PEARP : 338 stations

CEPMMT : 19 stations

PEARP : 486 stations

Bleu : CEPMMT meilleur (90% de certitude selon un test de ré-échantillonnage)Rouge : PEARP meilleur (90% de certitude)

Distribution par taille de bassin (BSS)

Q10 Jour 1

Q10 Jour 2

Q90 Jour 2

Q90 Jour 1

CEPMMT

PEARPTailles des bassins Tailles des bassins

Tailles des bassinsTailles des bassins

Variance d’erreur d’observations

Erreurs des mesures des stations indépendantes : matrice diagonale

Tests sur des cas synthétiques : 2e méthode meilleure (Nash) et donc retenue

Répartition spatiale de la variance d’erreur d’ébauche

Moyenne pondérée des 2 couches

Couche 3 uniquementCouche 2 uniquement

B et R diagonales

B estimée en perturbant l’analyse météorologique SAFRAN, puis comparaison de l’humidité obtenue avec l’humidité de référence

R estimée selon les débits observés

R et B calibrées grâce à un unique coefficient

Expériences jumelles

Variable d’état = moyenne pondérée des humidités des 2 couchesAssimilation sur une période de 3 mois, tous les 5 jours, fenêtre d’assimilation de 5 joursEtat initial modifié, obs = simulation de référence

Convergence assez rapide malgré les non-linéarités

Scores avec le BLUE itéré

Un exemple de l’impact sur les prévisions d’ensemble des débits

IS2

IS1

Sans assimilation

Ranked Probability Skill Score

RMSE par taille de bassin

Jour 1 Jour 10

Résolution

Fiabilité

Incertitude

Taux de fausses alarmes

Taux de réussite

Jour 1

Jour 10

Sans assimilationAvec EI1Avec EI2

« Improvement of ensemble streamflow predictions over France of the SAFRAN-ISBA-MODCOU model » Guillaume Thirel (CNRM-GAME/GMME/MOSAYC) PhD Director :

Documents

ensemble discharges

pearp eps slide

climatologie slide

data assimilation system

sim observations

discharges article rousset

epss ecmwf eps

discharges soil water