Date post: | 15-Jan-2016 |
Category: |
Documents |
View: | 213 times |
Download: | 0 times |
Page 1
The Construction and Use of Linear Models in Large-scale Data Assimilation
Tim Payne
Large-Scale Inverse Problems and Applications in the Earth Sciences
October 24th 2011
© Crown copyright 2011 Page 2Page 2
Part I.
The Construction of Linear Models in Data Assimilation
© Crown copyright 2011 Page 3Page 3
Notation
© Crown copyright 2011 Page 4Page 4
Update-Prediction Cycle
© Crown copyright 2011 Page 5Page 5
First Strategy – exact evolution of covariances
© Crown copyright 2011 Page 6Page 6
Second strategy – EKF using tangent-linear
© Crown copyright 2011 Page 7Page 7
Third strategy – EKF using best linear approximation
© Crown copyright 2011 Page 8Page 8
Explicit formula for best linear approximation
© Crown copyright 2011 Page 9Page 9
Basic properties of best linear approximation
© Crown copyright 2011 Page 10Page 10
Cloud function
T
Seek a rational basis on which to 'regularise' in 4D-Var the Smith cloud scheme of 1990,
where cloud fraction is expressed as a function of
( , )=
(1 ) ( , )
Cloud fraction and its de
sat LN
c sat L
C
q q T pQ
RH q T p
N
212
212
rivative w.r.t. Q are
0 1 0 1
(1 ) 1 0 (1 ) 1 0'
1 (1 ) 0 1 (1 ) 0 1
1 1 0 1
How should one regularise this function?
Met Office ha
N N
N N N N
N N N N
N N
for Q for Q
Q for Q Q for QC C
Q for Q Q for Q
for Q for Q
21N2d used regularisation [1 tanh(2 )] with derivativeC'=sech (2Q )NC Q
© Crown copyright 2011 Page 11Page 11
Smith Cloud Scheme with ‘ad hoc’ Regularisation
212
212
0 1
(1 ) 1 0
1 (1 ) 0 1
1 1
N
N N
N N
N
for Q
Q for QC
Q for Q
for Q
0 1
(1 ) 1 0'
(1 ) 0 1
0 1
N
N N
N N
N
for Q
Q for QC
Q for Q
for Q
12Standard regularisation [1 tanh(2 )]NC Q 2' sech (2 )NC Q
© Crown copyright 2011 Page 12Page 12
Smith Cloud Scheme with ‘Optimal’ Regularisation
1
. 1 Gaussian
~ (0, )
1 | |pdf ( ) exp
2
Eg p
Q N
xf x
| |
Opt reg LS:
E{C(Q+ )}
1( )exp( )
2x
Q
C Q x dx
2
| |3
E{C(Q+ ) Q}Opt reg PF:
1( ) exp( )
2x
Q
C Q x x dx
N N N N N
N N NN N N N N 2
N
Optimal regularisation: C(Q + Q ) F(Q )+T(Q ) Q
E[C(Q + Q ) Q ]If E( Q ) 0 then F(Q )=E[C(Q + Q )] and T(Q )=
E[ Q ]
© Crown copyright 2011 Page 13Page 13
Incremental 4D-Var
© Crown copyright 2011 Page 14Page 14
Options for linearisation step required for incremental 4D-Var
© Crown copyright 2011 Page 15Page 15
Gain matrix implied by each option
© Crown copyright 2011 Page 16Page 16
Advantage of BLA over TL in incremental 4D-Var
© Crown copyright 2011 Page 17Page 17
Pseudo Chain-rule for best linear approximation
© Crown copyright 2011 Page 18Page 18
Use of best linear approximation in EKF
© Crown copyright 2011 Page 19Page 19
The prior covariance implied by different approximations
© Crown copyright 2011 Page 20Page 20
Prior covariance using best linear estimate
© Crown copyright 2011 Page 21Page 21
Prior covariance using the best linear estimate always underestimates the true prior
© Crown copyright 2011 Page 22Page 22
The Duffing Map
© Crown copyright 2011 Page 23Page 23
100,000 iterates of Duffing Map
© Crown copyright 2011 Page 24Page 24
Reminder of EKF algorithm
© Crown copyright 2011 Page 25Page 25
Prior covariance for Duffing Map
© Crown copyright 2011 Page 26Page 26
Mean square analysis error in Duffing map: TL and best linear estimate compared
© Crown copyright 2011 Page 27Page 27
Part II.
The Use of Linear Models in Data Assimilation
© Crown copyright 2011 Page 28Page 28
Linearisation error in 4D-Var as used in real numerical weather prediction models
© Crown copyright 2011 Page 29Page 29
© Crown copyright 2011 Page 30Page 30
Linear model for evolution of increments
© Crown copyright 2011 Page 31Page 31
Linearisation error as a stochastic error
© Crown copyright 2011 Page 32Page 32
Issues in forming EKF
© Crown copyright 2011 Page 33Page 33
Signal model for system with time correlated linearisation error
© Crown copyright 2011 Page 34Page 34
EKF with time correlated linearisation error
© Crown copyright 2011 Page 35Page 35
Parameters for filter including linearistion error
© Crown copyright 2011 Page 36Page 36
Example: L95, nearly perfect full model, persistence for linear model
© Crown copyright 2011 Page 37Page 37
Example: L95, nearly perfect full model, persistence for linear model, results
© Crown copyright 2011 Page 38Page 38
Variational version: weak constraint 4D-Var allowing for time correlated linearisation error
© Crown copyright 2011 Page 39Page 39
Remarks on variational form
© Crown copyright 2011 Page 40Page 40
Long window weak constraint 4D-Var allowing for linearisation error, same example
© Crown copyright 2011 Page 41Page 41
Summary to Part I
© Crown copyright 2011 Page 42Page 42
Summary to Part II
© Crown copyright 2011 Page 43Page 43
The End