contributions / input by: hendrik reich, andreas rhodin, klaus stephan, werner wergen (dwd)
DESCRIPTION
PP Kenda : Status Report [email protected] Deutscher Wetterdienst, D-63067 Offenbach, Germany. Contributions / input by: Hendrik Reich, Andreas Rhodin, Klaus Stephan, Werner Wergen (DWD) Daniel Leuenberger, Tanja Weusthoff (MeteoSwiss) Marek Lazanowicz (IMGW) - PowerPoint PPT PresentationTRANSCRIPT
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
Contributions / input by:
Hendrik Reich, Andreas Rhodin, Klaus Stephan, Werner Wergen (DWD)
Daniel Leuenberger, Tanja Weusthoff (MeteoSwiss)
Marek Lazanowicz (IMGW)
Mikhail Tsyrulnikov (HMC)
PP Kenda : Status Report [email protected]
Deutscher Wetterdienst, D-63067 Offenbach, Germany
• status & outlook
• general issues in the convective scale experiments for assessing importance of km-scale details in IC
• deterministic analysis
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
Task 1: General issues in the convective scale and evaluation of COSMO-DE-EPS
Purpose: Guides decision how resources will be spent on/ split betw. LETKF and SIR
(COSMO-NWS and universities); part of the learning process
main disadvantage of LETKF: assumes Gaussian error distributions
Task 1.1.A: investigate non-Gaussianity by means of O – B statistics (convective / larger scales, different forecast lead times): provides an upper limit estimate of the non-Gaussianity to deal with
talk by Daniel Leuenberger:Statistical characteristics of high-resolution COSMO Ensemble forecastsin view of Data Assimilation
Task 1.2: investigate non-Gaussianity by examining perturbations of very-short range(2009) forecasts from COSMO-DE-EPS
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
RWWRP / THORPEX Workshop on 4D-Var and EnKF intercomparison• we should continue with KENDA (LETKF) as planned
(4DVAR / EnKF ok for HR DA: everybody pushes the way further on the current road; ensemble size stable (30 – 40); synergistic approaches)
• model errors a (the) key issue for advanced DA important for DA to work on model improvement
• worthwhile to obtain options for EnKF implementations which also make use of 3DVar(by implementing tasks 2.1 , 2.5)
talk by Breogan Gomez:Single-column experiments on the vertical localisation in the LETKF
difficulty of assimilating non-local satellite data and achieving good resolution in local analysis
Task 1.4: M. Tsyrulnikov: Review on Hunt et al. implementation of LETKF
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
Task 1.1 D: assess importance of km-scale details versus larger-scale conditions in the IC
(do we have to analyse the small scales, or is it sufficient to analyse the large scales, as e.g. incremental 4DVAR (ECMWF) would do ?)
Comparison: ‘IEU’: IC from interpolated COSMO-EU analysis of ass. cycle (no LHN)
‘IDE’: IC from COSMO-DE analysis of ass. cycle (no LHN)
‘LHN’: IC from COSMO-DE analysis of ass. cycle (IDE + LHN for use of radar-
derived precipitation)
Note: – IC from assimilation cycle → late cut-off, very similar set of observations
– identical correlation functions (scales) used in nudging for IEU and IDE
– identical soil moisture, taken from IEU (with variational soil moisture initialisation)
– model version as operational in summer 2009
Period: 31 May – 13 June 2007: air mass convection
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
00 UTC runs 06 UTC runs
ETS
FBI
LHNIDE (no-LHN)IEU (coarse IC)
31.05. – 13.06.07: air-mass convection
# radar ‘obs’ with rain
LHNIDE (no-LHN)IEU (coarse IC)
time of day time of day
0.1 mm
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
00 UTC runs 06 UTC runs
ETS
FBI
LHNIDE (no-LHN)IEU (coarse IC)
31.05. – 13.06.07: air-mass convection
# radar ‘obs’ with rain
LHNIDE (no-LHN)IEU (coarse IC)
time of day time of day
1 mm
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
12 UTC runs 18 UTC runs
ETS
FBILHNIDE (no-LHN)IEU (coarse IC)
31.05. – 13.06.07: air-mass convection
0.1 mm
# radar ‘obs’ with rain
LHNIDE (no-LHN)IEU (coarse IC)
time of day time of day
12 18
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
12 UTC runs 18 UTC runs
ETS
FBI
LHNIDE (no-LHN)IEU (coarse IC)
31.05. – 13.06.07: air-mass convection
1 mm
# radar ‘obs’ with rain
LHNIDE (no-LHN)IEU (coarse IC)
time of day time of day
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
12-UTC runs,6 – 18 h fcsts.
(nighttime precipitation)
radar (12-h sum) LHN (fine-scale IC)IEU (coarse IC)
10 June 18 UTC
– 11 June 06 UTC
2007
4 June 18 UTC
– 5 June 06 UTC
2007
Examples
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
Task 1.1.D: assess importance of km-scale details
Results of comparison to coarse-scale ‘IEU’ :
─ ‘IDE’ better than ‘IEU’ for 12- and 18-UTC runs up to +15h (similar FBI, higher ETS) similar for 0- and 6-UTC runs
─ improvement by LHN during first 6 hours
─ ‘LHN’ has better precipitation patterns (6 – 18 h forecast) than ‘IEU’ in some cases
RadarWith LHN Without LHN (dashed: determininistic)
Past experiments (Leuenberger):environment affects impact of fine-scaledetails in analysis from LHN
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
Task 2: Technical implementation of an ensemble data assimilation framework / LETKF
analysis step (LETKF) outside COSMO code ensemble of independent COSMO runs up to next analysis time separate analysis step code, LETKF included in 3DVAR code of DWD
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
perturbed forecasts
– ensemble mean forecast
(local, inflated) transform matrix Wa analysis mean perturbed
analyses
LOCAL Ensemble Transform Kalman Filter (LETKF)(chart based on a slide of Neil Bowler, UK MetOffice)
= background perturbations Xb
→ flow-dep. backgr. error covar. analysis error covariance
in the (5-dimensional) sub-space S spanned by background perturbations :(linearise obs operator around the mean of ensemble of simulations of an obs)
explicit solution for minimisation of cost function (Hunt et al., 2007)
( - ) + =( - ) + =
( - ) + =( - ) + =( - ) + =
0.9 Pert 1-0.1 Pert 2-0.1 Pert 3-0.1 Pert 4-0.1 Pert 5
aa xw
)( )()( ibibb H xyy
→ analysis ensemble member : )()()( iaabbiaaia WwXxXxx 1/2PW aw
a k )1(
and in model space : Tbaw
baabba XPXP,wXxx
-111 )1( bTbaw
bTbaw
a k YRY I P,yyRYPw O → analysis (mean) and analysis error covariance : bibb yyY )(:columns of
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
Analysis for a deterministic forecast run : use Kalman Gain K of analysis mean
11
1)1()(
RYYRY I XwXK,xyKxx O TbbTbbabBBA kH
K is already computed in the (L)ETKF ;
if the resolution of the deterministic run is higher than of ensemble, the analysis increments have to be interpolated to the fine grid
determ.
KXw1
ba,bX
Kalman gain / analysis increments not optimal, if deterministic run (strongly) deviates from analysis mean of ensemble
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
LETKF
COSMO
read obs (NetCDF) + Grib analysis of ensemble member
compute obs–fg (obs. increm.) + QC (contains obs operator)
write NetCDF feedback files (obs + obs–fg + QC flags) + Grib files (model)
read ensemble of NetCDF feedback files + ensemble of COSMO S-R forecast Grib files
perform LETKF (based on obs–fg values around each grid pt.,calc. transformation matrices and analysis (mean & pert.))(adapt: C-grid, specific variables (w), efficiency)
write ensemble of COSMO S-R analysis Grib files + NetCDF feedback files with additional QC flags ( verif.)
exp.system
Task 2.3: finished
Task 2.2: almost finishedTask 2.4: not yet done
However: scripts written to do a few stand-alone cycles with LETKF → preliminary tests can start soon
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
y: observationsx: model stated = y – H(x): innovationH: observation operatorqf: quality flagQC: quality control
Verification / diagnostics: new ‘stat’ tool : compute model (forecast) – obs , as input for verification
GRIBnew tool (included in 3DVar code)
NetCDFobservation y
qf
read obswrite
feed-
backxH(x)
QCH
read
feedback
y
qf
H(x)read grib
NetCDFfeedbacky, qf, d
want have capability of computing distance of model (ensemble member) to observationsthat have not been used previously in a COSMO run ( SIRF)
need to include
NetCDFobservation
QCH& in ‘stat’ module
common codes in COSMO & ‘stat’ tool
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
Task 2.1: extract from COSMO:
- ‘library 2’ : in progress
- ‘library 1’ : finished
Task 2.5: - in progress: include above libraries in 3DVar environment, (translate COSMO data structure into 3DVar data structure and vice versa)
- not yet started: extend flow control (e.g. reading several Grib files and temporal interpolat.)
Task 2: Technical implementation of an ensemble data assimilation framework / LETKF
for verification: ‘stat’-module: compute model (forecast) – obs :adapt verification mode of 3DVar/LETKF package
Advantages:
– COSMO obs operators available in 3DVAR/LETKF environment 3DVar/ EnKF approaches requiring 3DVar in principle applicable to COSMO LETKF for ICON will require COSMO obs operators in the future
– 1 common code for GME/ICON and COSMO to produce input for diagnostics / verif..
Disadvantages: – more complex code for this diagnostic task
– possibly additional transformation from COSMO data structure into 3DVAR data structure and vice versa required for new COSMO obs operators.
observation operators H with QC
modules for reading obs from NetCDF
Task 2.6: Adapt ensemble-related diagnostic tools : not yet started
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
Task 3: Evaluate and tune LETKF
address primary scientific issues related to the LETKF on the convective scale(using only in-situ data at first)
• investigate / control noise, adapt in view of non-hydrostatic aspectsif mixing of hydrostatically balanced model states (ensemble forecasts) varies with height (vertical localisation), the resulting model analysis will usually not be hydrostatically balanced
→ get version that can be used to do first tests with radar radial winds
• primary scientific issues:– model model (physical) perturbations– multiplicative covariance inflation– localisation (multi-scale DA)– additive covariance inflation with red noise, backscatter, statistical forecast error
covariances (‘3DVAR-B’) ‘– bias removal– noise control (update frequency, digital filter initialisation, lateral boundary cond.)– (convection initiation (warm bubbles, LHN)
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
Task 4: Inclusion of additional observations
Research funded by DWD to develop suitably sophisticated + efficient observation operators forradar data (radial winds + (3-dim.) reflectivity) (Blahak + 2 PhD’s)
Task 4.1: Radar radial winds:• implement simple obs operator, then test and tune• particular issues: obs error variances and correlations, superobbing, thinning, localisation
Task 4.2: Radar reflectivity (after 2010):
Task 4.3: Ground-based GPS slant path delay (after 2010):• particular issue: localisation for non-local obs
Task 4.4: Cloud info based on satellite and conventional data(DWD: applied for Eumetsat fellowship, start end of 2010)
• derive incomplete cloud analysis, use obs increments of cloud or humidity• use SEVIRI brightness temperature directly in LETKF in cloudy (+ cloud-free) conditions
Task 5: Sequential importance resampling (SIR) filter
research projects continue within DAQUA (MeteoSwiss investigating on norms)
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
thank you for your attention
KENDA (Km-Scale Ensemble-based Data Assimilation)
COSMO General Meeting, Offenbach, 7 – 11 Sept. 2009KENDA [email protected]
set up (G-) SIRF test with COPS period
time
Weighting +
Resam
pling
Guiding steps
Weighting +
Resam
pling
final weightingafter free forecast is computed
SREPS (ass.) SREPS (pred.)
radar / satellite / in-situ obs.
free forecast
… …
HRES
select bestmembers( ≤10 )
Set up standard and G-SIRF with and without standard data assimilation (MIUB, DWD)
• assess impact of conventional DA (LHN, PIB) on ensemble development (spread generation, keeping ensemble on track)
• implement optimal stepping to a new driving mesoscale ensemble
Evaluate classical and spatial (object oriented, fuzzy) metricsfor weighting mesoscale (SREPS) and km-scaleensemble members (DLR, MCH)• assess correlation of metrics betw. models of different res.• assess persistence of skill in different metrics