Deutscher Wetterdienst
COSMO-DE-EPS
Susanne Theis, Christoph Gebhardt, Michael Buchhold,
Zied Ben Bouallègue, Roland Ohl, Marcus Paulat, Carlos Peralta
with support by: Helmut Frank, Thomas Hanisch, Ulrich Schättler, etc
COSMO GM – September 2010
NWP Model COSMO-DE
grid size 2.8 km
without parametrization
of deep convection
(convection-permitting)
lead time 0-21 hours
operational since April 2007
COSMO-EU
GME
COSMO-DE
COSMO GM – September 2010
Plans for a COSMO-DE Ensemble
How many ensemble members?
preoperational: 20 members
operational: 40 members
When?
preoperational: 2010
operational: 2012
COSMO GM – September 2010
COSMO-DE-EPS production steps
„variations“ withinforecast system
Ensemble products:
- mean- spread- probabilities- quantiles- ...
ensemble members
COSMO GM – September 2010
COSMO-DE-EPS production steps
„variations“ withinforecast system
Ensemble products:
- mean- spread- probabilities- quantiles- ...
ensemble members
+ verification
+ postprocessing
COSMO GM – September 2010
COSMO-DE-EPS production steps
„variations“ withinforecast system
Ensemble products:
- mean- spread- probabilities- quantiles- ...
ensemble members
11
next slides:step 1, generation of members
next slides:step 1, generation of members
COSMO GM – September 2010
Generation of Ensemble Members
Variations in Forecast System
for the Representation of Forecast Uncertainty
Initial Conditions Boundaries Model Physics
COSMO GM – September 2010
Generation of Ensemble Members
“multi-model”
driven by different global models
Variations in Forecast System
for the Representation of Forecast Uncertainty
Initial Conditions Boundaries Model Physics
COSMO GM – September 2010
Generation of Ensemble Members
“multi-model”
driven by different global models
“multi-model”
COSMO-DE initial conditions modified by different global models
Variations in Forecast System
for the Representation of Forecast Uncertainty
Initial Conditions Boundaries Model Physics
COSMO GM – September 2010
Generation of Ensemble Members
“multi-configuration”
different configurations of COSMO-DE model
Variations in Forecast System
for the Representation of Forecast Uncertainty
Initial Conditions Boundaries Model Physics
“multi-model”
driven by different global models
“multi-model”
COSMO-DE initial conditions modified by different global models
COSMO GM – September 2010
Generation of Ensemble Members
The Ensemble Chain
COSMO-DE 2.8km
COSMO 7km
globalmm/24h
COSMO-DE
COSMO GM – September 2010
Generation of Ensemble Members
The Ensemble Chain
COSMO-DE 2.8km
COSMO 7km
globalmm/24h
COSMO-DE
plus the variations of• initial conditions• model physics
COSMO GM – September 2010
Generation of Ensemble Members
Which computers are used?
at ECMWF: „7 km Ensemble“
at DWD: COSMO-DE-EPS
COSMO-DE 2.8km
COSMO 7km
global
ECMWF DWD
COSMO GM – September 2010
Generation of Ensemble Members
GFSIFSGME
COSMO 7km
…etc…
transfer of data
Which computers are used?
at ECMWF: „7 km Ensemble“
at DWD: COSMO-DE-EPS
COSMO GM – September 2010
Generation of Ensemble Members
GFSIFSGME
COSMO 7km
…etc…
Which computers are used?
at ECMWF: „7 km Ensemble“
at DWD: COSMO-DE-EPS
transfer of data
Status: in testing phase
(so far: COSMO-SREPS)
COSMO GM – September 2010
variation of initial conditions
Generation of Ensemble Members
COSMO GM – September 2010
variation of initial conditions
global forecasts
Generation of Ensemble Members
COSMO GM – September 2010
variation of initial conditions
COSMO-DE 2.8 km
global forecasts
COSMO 7 km
ic
Generation of Ensemble Members
COSMO GM – September 2010
variation of initial conditions
COSMO-DE 2.8 km
COSMO-DEassimilation
global forecasts
COSMO 7 km
IC
ic
Generation of Ensemble Members
COSMO GM – September 2010
variation of initial conditions
modify initial conditions of COSMO-DE
by using differences between
the COSMO 7km initial conditions
IC´ = F (IC, ic – icref)
COSMO-DE 2.8 km
COSMO-DEassimilation
global forecasts
COSMO 7 km
IC
ic
Generation of Ensemble Members
COSMO GM – September 2010
variation of „model physics“
Generation of Ensemble Members
COSMO GM – September 2010
variation of „model physics“
11
22
33
55
44
entr_sc
rlam_heat
rlam_heat
q_crit
tur_len
different configurationsof COSMO-DE 2.8 km:
Generation of Ensemble Members
COSMO GM – September 2010
variation of „model physics“
selection of configurations:
subjective,
based on experts and verification
selection criteria:
1. large effect on forecasts
2. no „inferior“ configuration
11
22
33
55
44
entr_sc
rlam_heat
rlam_heat
q_crit
tur_len
different configurationsof COSMO-DE 2.8 km:
Generation of Ensemble Members
COSMO GM – September 2010
future changes
- extension to 40 members
- switch to ICON as driving ensemble
(model ICON currently under development)
- apply an Ensemble Kalman Filter for initial condition perturbations
(EnKF currently under development for data assimilation)
Generation of Ensemble Members
COSMO GM – September 2010
COSMO-DE-EPS production steps
„variations“ withinforecast system
Ensemble products:
- mean- spread- probabilities- quantiles- ...
ensemble members
22
next slides:step 2, generating „products“
next slides:step 2, generating „products“
COSMO GM – September 2010
Generation of „Ensemble Products“
variables (list will be extended):
1h-precipitation
wind gusts
2m-temperature
ensemble „products“:
probabilities
quantiles
ensemble mean
min, max
spread
ensemble products:
- mean- spread- probabilities- quantiles- ... GRIB2
22
GRIB1
COSMO GM – September 2010
further improvement:
adding a spatial neighbourhood
adding simulations started a few hours earlier
Generation of „Ensemble Products“
COSMO GM – September 2010
further improvement:
adding a spatial neighbourhood
adding simulations started a few hours earlier
additional product:
probabilities
with upscaling
%
event somewherein 2.8 km Box
event somewherein 28 km Box
Generation of „Ensemble Products“
COSMO GM – September 2010
COSMO-DE-EPS production steps
„variations“ withinforecast system
Ensemble products:
- mean- spread- probabilities- quantiles- ...
ensemble members
33
next slides:step 3, visualization in NinJo
next slides:step 3, visualization in NinJo
COSMO GM – September 2010
Visualization in NinJo
new development: „Ensemble Layer“
for NinJo version 1.3.6
released in 2010
COSMO GM – September 2010
COSMO-DE-EPS production steps
„variations“ withinforecast system
Ensemble products:
- mean- spread- probabilities- quantiles- ...
ensemble members
COSMO GM – September 2010
COSMO-DE-EPS production steps
„variations“ withinforecast system
Ensemble products:
- mean- spread- probabilities- quantiles- ...
ensemble members
+ verification
+ postprocessing
COSMO GM – September 2010
Verification Results
GEBHARDT, C., S.E. THEIS, M. PAULAT, Z. BEN BOUALLÈGUE, 2010:
Uncertainties in COSMO-DE precipitation forecasts introduced by model perturbationsand variation of lateral boundaries. Submitted to Atmospheric Research.
very first aim:
Does the ensemble meet some basic requirements?
results:
- ensemble spread is present
- members are of similar quality
- ensemble is superior to individual forecasts
COSMO GM – September 2010
Postprocessing / Calibration
„variations“ withinforecast system
Ensemble products:
- mean- spread- probabilities- quantiles- ...
ensemble members
COSMO GM – September 2010
Postprocessing / Calibration
„variations“ withinforecast system
Ensemble products:
- mean- spread- probabilities- quantiles- ...
ensemble members
COSMO GM – September 2010
Motivation for Postprocessing / Calibration
Aim: improve the quality
learn from past forecast errors
derive statistical connections
apply them to real-time ensemble forecasts
Forecast Obs
real-timeforecasts
historical data
COSMO GM – September 2010
Methods for Postprocessing / Calibration
First Approach:
logistic regression
ensemble „products“:
- mean- spread- probabilities- quantiles- ...
statisticalpostprocessing
COSMO GM – September 2010
Methods for Postprocessing / Calibration
ensemble „products“:
- mean- spread- probabilities- quantiles- ...
statisticalpostprocessing
First Approach:
logistic regression
Plan: preoperational in 2011
for precipitation
COSMO GM – September 2010
Research for Postprocessing / Calibration
in addition:
Research at Universities, funded by DWD
- University of Bonn:
Petra Friederichs, Sabrina Bentzien
Methods: Quantile Regression, Extreme Value Statistics
- University of Heidelberg:
Tilmann Gneiting, Michael Scheuerer
Methods: Bayesian Model Averaging, Geostatistics
COSMO GM – September 2010
COSMO-DE-EPS production steps
„variations“ withinforecast system
Ensemble products:
- mean- spread- probabilities- quantiles- ...
ensemble members
+ verification
+ postprocessing
COSMO GM – September 2010
Plans COSMO-DE-EPS
2010: start of preoperational phase
(20 members)
2010-2012: further extensions
statistical postprocessing
40 members
2012: start of operational phase
convection-permitting ensemble operation convection-permitting ensemble operation