Upload
domien
View
215
Download
0
Embed Size (px)
Citation preview
Towards a longer assimilation window in 4D-Var
Yannick Tremolet
Thanks to Paul Poli
ECMWF
October 2011
Y. Tremolet Long window 4D-Var October 2011
Outline
1 What we want to do
2 What we can do
3 ResultsModel Error Aspects24h Window: Operational System24h Window: Re-analysis System
4 Final Comments
Y. Tremolet Long window 4D-Var October 2011 1 / 21
Outline
1 What we want to do
2 What we can do
3 ResultsModel Error Aspects24h Window: Operational System24h Window: Re-analysis System
4 Final Comments
Y. Tremolet Long window 4D-Var October 2011
Weak Constraint 4D-Var
For Gaussian, temporally-uncorrelated model error, the weak constraint4D-Var cost function is:
J(x) =1
2(x0 − xb)TB−1(x0 − xb)
+1
2
n∑i=0
[Hi (xi )− yi ]TR−1
i [Hi (xi )− yi ]
+1
2
n∑i=1
[xi −Mi (xi−1)]TQ−1i [xi −Mi (xi−1)]
Do not reduce the control variable using the model and retain the 4D natureof the control variable.
Account for the fact that the model contains some information but is notexact by adding a model error term to the cost function.
This problem can be solved in parallel (saddle-point algorithm, no need forinverse of covariances, preconditioning is being investigated).
Y. Tremolet Long window 4D-Var October 2011 2 / 21
Longer is better
Theory says: long window weak constraint 4D-Var is equivalent to a full rankKalman smoother (Fisher et al., 2005, Menard and Daley, 1996).
Long window weak constraint 4D-Var works for simple systems (Lorenz 95,QG):
0 12 24 36 48 60 72 84 96 108 120 132 144 156 1680.4
0.45
0.5
0.55
0.6
0.65
0.7
0.75Mean Analysis and First−Guess Error for Different Window Lengths
Length of the Assimilation Window (hours)
RM
S E
rror
for
Non
−di
men
sion
al S
trea
mfu
nctio
n
Initial quess using scaled qbarAnalysis using scaled qbar
H. Auvinen and M. Fisher
Y. Tremolet Long window 4D-Var October 2011 3 / 21
Long Window Weak Constraint 4D-Var
�������
�� ��
�
�������
��� �
����
(1) Weak constraint 4D-Var (2) Extended window
�������
�� ��
� ��
�������
��
� ���
(3) Initial term has converged (4) Assimilation window is moved forward
This implementation is an approximation of weak contraint 4D-Var with anassimilation window that extends indefinitely in the past...
...which is equivalent to a (full rank) Kalman smoother that has been runningindefinitely.
And B is a problem of the past! Only the error characteristics of thefundamental ingredients of the DA problem remain.
Y. Tremolet Long window 4D-Var October 2011 4 / 21
Outline
1 What we want to do
2 What we can do
3 ResultsModel Error Aspects24h Window: Operational System24h Window: Re-analysis System
4 Final Comments
Y. Tremolet Long window 4D-Var October 2011
4D-Var with Model Error Forcing
In practice, weak constraint 4D-Var is still difficult to implement (in the IFS).
Change of variable:
J(x0, η) =1
2
n∑i=0
[H(xi )− yi ]TR−1
i [H(xi )− yi ]
+1
2(x0 − xb)TB−1(x0 − xb) +
1
2
n∑i=1
ηTi Q−1
i ηi
with xi =Mi (xi−1) + ηi
ηi represents model error in a time step,
ηi has the same dimension as a 3D state.
Y. Tremolet Long window 4D-Var October 2011 5 / 21
4D-Var with Constant Model Error Forcing
Approximation: model error is constant.
J(x0, η) =1
2
n∑i=0
[H(xi )− yi ]TR−1
i [H(xi )− yi ]
+1
2(x0 − xb)TB−1(x0 − xb) +
1
2ηTQ−1η
with xi =Mi (xi−1) + η
η represents model error in a time step,
η has the same dimension as a 3D state.
The number of degrees of freedom doubles.
Y. Tremolet Long window 4D-Var October 2011 6 / 21
Weak Constraints 4D-Var for Systematic Model Error
For random model error, the 4D-Var cost function is:
J(x0, η) =1
2
n∑i=0
[H(xi )− yi ]TR−1
i [H(xi )− yi ]
+1
2(x0 − xb)TB−1(x0 − xb) +
1
2ηTQ−1η
For systematic model error:
J(x0, η) =1
2
n∑i=0
[H(xi )− yi ]TR−1
i [H(xi )− yi ]
+1
2(x0 − xb)TB−1(x0 − xb) +
1
2(η − ηb)TQ−1(η − ηb)
Test case: model bias in the stratosphere.
Y. Tremolet Long window 4D-Var October 2011 7 / 21
Model Error Covariance Matrix
Currently, tendency differences between integrations of the members of anensemble are used as a proxy for samples of model error.
Statistics of model drift (for systematic model error).
Use results from stochastic representation of uncertainties in EPS.
It is possible to derive an estimate of HQHT from cross-covariances betweenobservation departures produced from pairs of analyses with different lengthwindows (R. Todling).
Is it possible to extract model error information using the relationPf = MPaMT + Q?
Model error is correlated in time: Q should account for time correlations.How?
How to account for flow dependence?
Y. Tremolet Long window 4D-Var October 2011 8 / 21
Outline
1 What we want to do
2 What we can do
3 ResultsModel Error Aspects24h Window: Operational System24h Window: Re-analysis System
4 Final Comments
Y. Tremolet Long window 4D-Var October 2011
Outline
1 What we want to do
2 What we can do
3 ResultsModel Error Aspects24h Window: Operational System24h Window: Re-analysis System
4 Final Comments
Y. Tremolet Long window 4D-Var October 2011
Weak Constraints 4D-Var with Cycling Term
01 15Jun
01 15Jul
01 15Aug
01 15Sep
01 15Oct
01 15Nov
01 15Dec
01 15Jun
01 15Jul
01 15Aug
01 15Sep
Weak constraints 4D-Var with cycling – MetOp-A AMSU-A Tb 13 N. Hemis – Model level 14
Weak constraints – MetOp-A AMSU-A Tb 13 N. Hemis – Model level 14
01 15Oct
01 15Nov
01 15Dec
0.6
0.4
0.2
0.0
–0.2
–0.4
–0.6
ΔT
(K
)
0.6
0.4
0.2
0.0
–0.2
–0.4
–0.6
ΔT
(K
)
OBS-BG OBS-AN OBS Bias Increment Model Error
The short term forecast is improved with the model error cycling.Weak constraints 4D-Var can correct for seasonal bias (partially).
Y. Tremolet Long window 4D-Var October 2011 9 / 21
Observation Error or Model Error?
05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20
Oct
Nov
Dec
Jan
Feb
−0.5
0.0
0.5
1.0
1.5
2.0
2.5
�T (K)Weak constraints 4D-Var with cycling - Metop-A AMSU-A Tb 13 N.Hemis - Model level 14
OB-FGOB-ANObs BiasIncrementModel Error
05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20
Oct
Nov
Dec
Jan
Feb
−0.5
0.0
0.5
1.0
1.5
2.0
2.5
�T (K)
CY35R2 - Metop-A AMSU-A Tb 13 N.Hemis - Model level 14
OB-FGOB-ANObs BiasIncrement
Observation error bias correction can compensate for model error.
Y. Tremolet Long window 4D-Var October 2011 10 / 21
Weak Constraint 4D-Var
Model Error (K/day) Model Drift (K/day)
90N/ 90S/ MetgraF-3.0-2.00-1.80-1.6-1.40-1.20-1.00-0.8-0.60-0.4-0.20-0.10.100.20.400.60.801.001.201.41.601.802.0
-1.8-1.4
-1.4
-1.0
-1.0
-0.6
-0.6
-0.6
-0.6
-0.2
-0.2
-0.2
-0.2
-0.2
-0.2
-0.2
-0.2
0.1
0.1
0.1
0.1
0.1
0.1 0.1
0.1
0.1
0.1
0.4
0.4
0.4
0.4
0.40.4
0.4
0.4 0.40.4 0.4
0.8
0.8
0.8
0.8
1.2
1.2
1.2
1.2
1.6
1000mb 850mb 700mb
500mb
300mb
200mb
150mb
100mb
70mb
50mb
30mb
20mb
10mb
5mb
2mb
1.00mb
0.50mb
0.20mb
0.10mb
0.05mb
0.02mb
0.01mb 1
2
3
4
5
6
7
8
9
10
11
12
13
141516171819202122
24
2628303234363840424446485052545658606264687276808488
90N/ 90S/ MetgraF-3.0-2.0-1.5-1.0-0.70-0.5-0.30-0.20-0.15-0.10-0.07-0.04-0.02-0.010.010.020.040.070.100.150.200.30.500.71.01.52.0
-0.3
-0.3
-0.3
-0.2
-0.2
-0.2 -0.2
-0.2
-0.1
-0.1
-0.1
-0.1
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.00.0 0.0 0.0
0.0
0.0
0.0
0.0
0.0 0.0
0.0
0.0
0.0 0.0
0.1
0.1
0.10.1
0.1
0.10.1
0.1
0.2
0.2
0.2
0.2
0.2
0.2
0.20.2
0.2
0.5
1000mb 850mb 700mb
500mb
300mb
200mb
150mb
100mb
70mb
50mb
30mb
20mb
10mb
5mb
2mb
1.00mb
0.50mb
0.20mb
0.10mb
0.05mb
0.02mb
0.01mb 1
2
3
4
5
6
7
8
9
10
11
12
13
141516171819202122
24
2628303234363840424446485052545658606264687276808488
Temperature zonal means, December 2010
Y. Tremolet Long window 4D-Var October 2011 11 / 21
Weak Constraint 4D-Var
Mean (K/day) Standard Deviation (K/day)
90N/ 90S/ MetgraF-3.0-2.00-1.80-1.6-1.40-1.20-1.00-0.8-0.60-0.4-0.20-0.10.100.20.400.60.801.001.201.41.601.802.0
-1.8-1.4
-1.4
-1.0
-1.0
-0.6
-0.6
-0.6
-0.6
-0.2
-0.2
-0.2
-0.2
-0.2
-0.2
-0.2
-0.2
0.1
0.1
0.1
0.1
0.1
0.1 0.1
0.1
0.1
0.1
0.4
0.4
0.4
0.4
0.40.4
0.4
0.4 0.40.4 0.4
0.8
0.8
0.8
0.8
1.2
1.2
1.2
1.2
1.6
1000mb 850mb 700mb
500mb
300mb
200mb
150mb
100mb
70mb
50mb
30mb
20mb
10mb
5mb
2mb
1.00mb
0.50mb
0.20mb
0.10mb
0.05mb
0.02mb
0.01mb 1
2
3
4
5
6
7
8
9
10
11
12
13
141516171819202122
24
2628303234363840424446485052545658606264687276808488
90N/ 90S/ MetgraF0.000.050.100.150.200.250.300.350.400.450.500.600.700.8
0.1
0.1
0.1
0.1
0.1 0.1
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.3
0.3
0.30.3
0.30.3
0.4
0.40.
5
0.7
1000mb 850mb 700mb
500mb
300mb
200mb
150mb
100mb
70mb
50mb
30mb
20mb
10mb
5mb
2mb
1.00mb
0.50mb
0.20mb
0.10mb
0.05mb
0.02mb
0.01mb 1
2
3
4
5
6
7
8
9
10
11
12
13
141516171819202122
24
2628303234363840424446485052545658606264687276808488
Temperature zonal means, December 2010
Model error estimates vary rapidly in NH stratosphere.
Y. Tremolet Long window 4D-Var October 2011 12 / 21
Outline
1 What we want to do
2 What we can do
3 ResultsModel Error Aspects24h Window: Operational System24h Window: Re-analysis System
4 Final Comments
Y. Tremolet Long window 4D-Var October 2011
24h 4D-Var: Forecast Scores
-5
-4
-3
-2
-1
0
1
2
3
4
5
0 1 2 3 4 5 6 7 8 9 10 11
Forecast Day
00UTC | Confidence: 95.0 | Population: 42
Date: 20101120 00UTC to 20101231 00UTCN Hem Extratrop (lat 20.0 to 90.0, lon -180.0 to 180.0)
Correlation coefficent of forecast anomaly500hPa geopotential Overlaping 24h 4D-Var minus 12h 4D-Var
-5
-4
-3
-2
-1
0
1
2
3
4
5
6
0 1 2 3 4 5 6 7 8 9 10 11
Forecast Day
00UTC | Confidence: 95.0 | Population: 42
Date: 20101120 00UTC to 20101231 00UTCS Hem Extratrop (lat -90.0 to -20.0, lon -180.0 to 180.0)
Correlation coefficent of forecast anomaly500hPa geopotential Overlaping 24h 4D-Var minus 12h 4D-Var
Forecast scores for overlapping 24h 4D-Var with respect to 12h 4D-Var.
Y. Tremolet Long window 4D-Var October 2011 13 / 21
24h 4D-Var: Forecast Scores
0 1 2 3 4 5Forecast Range (days)
0
10
20
30
40
50
RM
SE
rror
(m)
12h 4D-Var 24h fc1 an1 24h fc1 an2 24h fc2 an2
With overlapping analysis windows, there are several analyses to start theforecast from and to verify against!
Warning: too few cases to draw conclusions from this figure.
Y. Tremolet Long window 4D-Var October 2011 15 / 21
24h 4D-Var: Observation Statisticsexp: fk5q / fk5q / DA (black) v. �ut / DA 2010120100 – 2010121512(12)
AIREP-T N. Hemisphere used T
nobsexp
132294
241448
257268
238009
135873
127386
395932
277642
234
0
+4374
+8280
+10298
+9255
+5374
+4798
+14242
+10116
+7
1000
850
700
500
400
300
250
200
150
100
1000
850
700
500
400
300
250
200
150
100
exp-refStandard deviation
Pre
ssu
re (
hP
a)
Pre
ssu
re (
hP
a)
Bias
Background departure o-b (ref ) Background departure o-b
Analysis departure o-a (ref ) Analysis departure o-a
0 0.2–0.2–0.4–0.6–0.8–1 0.4 0.6 0.8 11.81.20.60 2.4 3
exp: fk5q / fk5q / DA (black) v. �ut / DA 2010120100 – 2010121512(12)
TEMP-T N. Hemisphere used T
nobsexp
325105503663501548164051937345328653813142106420623561437972353233129726934
2403
+82+929
+1813+1653+1196+1068
+937+1053+1240+1362+1197+1285+1122
+983+849
+71
1000850700500400300250200150100
7050302010
5
Pre
ssu
re (
hP
a)
Pre
ssu
re (
hP
a)
100085070050040030025020015010070503020105
exp-refStandard deviation Bias
Background departure o-b (ref ) Background departure o-b
Analysis departure o-a (ref ) Analysis departure o-a
0 0.2–0.2–0.4–0.6–0.8–1 0.4 0.6 0.8 12.41.60.80 3.2 4
Y. Tremolet Long window 4D-Var October 2011 16 / 21
Outline
1 What we want to do
2 What we can do
3 ResultsModel Error Aspects24h Window: Operational System24h Window: Re-analysis System
4 Final Comments
Y. Tremolet Long window 4D-Var October 2011
Ps-only Re-analysis
Background and Analysis fit to Observations2004-07-01 to 2005-04-09
0 3 6 9 12 15 18 21 24Time (h)
0.5
1.0
1.5
2.0
Ps
obse
rvat
ion
fit(h
Pa)
Overlapping 24h 4D-Var 24h 4D-Var 12h 4D-Var
0 3 6 9 12 15 18 21 24Time (h)
0.5
1.0
1.5
2.0
Ps
obse
rvat
ion
fit(h
Pa)
Overlapping 24h 4D-Var 24h 4D-Var 12h 4D-Var
Y. Tremolet Long window 4D-Var October 2011 17 / 21
Ps-only Re-analysis
Forecast scores vs. operational analysisZ500, NH, 2004-07-01 to 2005-04-09
0 1 2 3 4 5 6 7 8 9 10Forecast Range (days)
20
30
40
50
60
70
80
90
100
Ano
mal
yC
orre
latio
n(%
)
Overlapping 24h 4D-Var 24h 4D-Var 12h 4D-Var
0 1 2 3 4 5 6 7 8 9 10Forecast Range (days)
20
30
40
50
60
70
80
90
100
Ano
mal
yC
orre
latio
n(%
)
Overlapping 24h 4D-Var 24h 4D-Var 12h 4D-Var
Y. Tremolet Long window 4D-Var October 2011 18 / 21
Ps-only Re-analysis
Verification against independent (unused) observations:I confirms positive results with overlapping windows,I shows that 24h 4D-Var without overlap is slightly better than 12h 4D-Var.
24h 4D-Var system has not been tuned.I Results should improve.
Why is 24h 4D-Var better in Ps-only re-analysis context?I Model error is small relative to other errors,I Kalman smoother rather than Kalman filter (in part),I Not enough observations to fully constrain the analysis in 12h 4D-Var,I Full observing system constrains the analysis so tightly that the assimilation
algorithm is not as important.
Y. Tremolet Long window 4D-Var October 2011 19 / 21
Outline
1 What we want to do
2 What we can do
3 ResultsModel Error Aspects24h Window: Operational System24h Window: Re-analysis System
4 Final Comments
Y. Tremolet Long window 4D-Var October 2011
24h Weak Constraint 4D-Var
In the current formulation of weak constraints 4D-Var (model error forcing):I Background term to address systematic error,I 24h assimilation window.
Observation biases can be an issue.I Experiment with bias corrected aircraft observations is starting.
Investigate physical meaning of model error estimates.I For the first time, we might be looking at model error!
Weak Constraints 4D-Var requires better knowledge of the statisticalproperties of model error.
Very good results in Ps-only experiments (re-analysis).
Kalman smoother is better at least for re-analysis.
Y. Tremolet Long window 4D-Var October 2011 20 / 21
Long Window Weak Constraints 4D-Var
Weak constraint 4D-Var with a 4D state control variable:I Four dimensional problem with a coupling term between sub-windows is a
smoother over the whole assimilation period.
Practical implementation is very difficult in current ECMWF system (code,scripts, archiving...).
We are re-designing our data assimilation system to make it all possible:Object Oriented Prediction System (OOPS).
I High level algorithms in C++,I Improved scalability, reliability, flexibility,I New algorithms are implemented (saddle point).
Y. Tremolet Long window 4D-Var October 2011 21 / 21