Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1...
Transcript of Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1...
![Page 1: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/1.jpg)
Ensemble Kalman Filter: Handling nonlinearity with the Running in Place (RIP) and Quasi Outer-‐Loop (QOL)
methods Shu-‐Chih Yang1
Eugenia Kalnay2 and Brian Hunt2
1 Na9onal Central University, Taiwan 2 University of Maryland
Special THANKS to Kayo Ide, Takemasa Miyoshi, Fuqing Zhang, Tamara Singleton and the UMD weather Chaos group
![Page 2: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/2.jpg)
Outline • Background
– Schemes of Running-‐in-‐Place (RIP)/Quasi-‐Outer Loop (QOL)
• Results with the Lorenz63 model
• Comparisons between itera&ve EnKFs
• Applica9ons of LETKF-‐RIP to regional data assimila9on
• Summary/Discussions
![Page 3: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/3.jpg)
• The Kalman Filter assump9on that the ensemble forecast perturba9ons are Gaussian is not valid if there is nonlinear growth • Nonlinearity depends on model dynamics, observaDons (accuracy, operators, sampling frequency) and model error
• Being able to use the nonlinear operators (M and H), EnKF handles some nonlinearity.
• Nonlinearity will increase the difficulty of the data assimila9on, par9cularly for EnKF.
Issues of nonlineariDes in data assimilaDon
![Page 4: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/4.jpg)
EnKF does not handle well long windows because ensemble perturbations become non-Gaussian. 4D-Var simply iterates and produces a more accurate analysis.
Analysis window (Days)
RMS error
LETKF 4D-‐Var
Average Analysis RMS error
![Page 5: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/5.jpg)
• Nonlinearity will increase the difficulty of the data assimila9on, par9cularly for EnKF.
• A disadvantage of ensemble-‐based KF is that ensemble perturba9ons become non-‐Gaussian under strong nonlinearity, and therefore needs short assimila9on windows.
• 4D-‐Var is a smoother: it keeps itera9ng un9l it fits the observa9ons within the assimila9on window as well as possible.
• EnKF doesn’t have the important outer loop as in the incremental 3D-‐Var and 4D-‐Var, widely used in opera9onal centers (ECMWF, NCEP, GMAO…)
Issues of nonlineariDes in data assimilaDon
![Page 6: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/6.jpg)
T799L91
T95L91 T159L91 T255L91
Figure from ECMWF, Anderson, 2005
Adjustments for the background trajectory and sensitivity matrix related to the linearization of the observation operator
![Page 7: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/7.jpg)
Dealing with nonlineari9es within EnKF
• EnKF is a sequen9al data assimila9on system where, ader the new data is used at the analysis 9me, it should be discarded. Only if the previous analysis and the new background are the most likely states given the past observaDons.
• For cases with strong nonlinear growth (e.g. the EnKF spin-‐up or the sudden change of the background dynamics), background ensemble can’t represent the state uncertainty and the most likely state is unlikely to happen!! −Filter divergence can take place.
![Page 8: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/8.jpg)
Filter divergence: when the trajectory is about to change regime
truth LETKF
Trajectory of x
perturbaDon grow rate
€
g =1nln
δxδx0
⎛
⎝ ⎜
⎞
⎠ ⎟
![Page 9: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/9.jpg)
Nonlinearity vs. Non-‐Gaussianity in EnKF
Truthh ETKFh
Trajectory of x
� truth, � observa9on, * analysis ensemble , * background ensemble
• Nonlinearity will distort the ensemble distribu9on and make it less Gaussian
• With non-‐Gaussian ensemble, the background ensemble quickly degrades.
• Sampled error sta9s9c lost track of the true dynamics
T1: Gaussian T2 :Non-‐Gaussian t3:Non-‐Gaussian
![Page 10: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/10.jpg)
Dealing with nonlineari9es within EnKF
• EnKF is a sequen9al data assimila9on system where, ader the new data is used at the analysis 9me, it should be discarded. Only if the previous analysis and the new background are the most likely states given the past observaDons.
• For cases with strong nonlinear growth (e.g. the EnKF spin-‐up or the sudden change of the background dynamics), background ensemble can’t represent the state uncertainty and the most likely state is unlikely to happen!!
• During strong nonlinearity, we wish to increase the influence of observaDons and should use the observaDons more than once if we can extract more informaDon.
![Page 11: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/11.jpg)
Kalman Filter and RIP with linear dynamics
• RIP is an algorithm that uses the same observation multiple times • Using RIP produces the same analysis means as the optimal KF. • The estimated error from KF-RIP with re-using the observations N times is the same as the one that would be computed from KF with higher observation accuracy, i.e., with an observation error variance divided by N.
![Page 12: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/12.jpg)
A linear model for state and error variance
KF vs. KF-‐RIP with a linear model
Using RIP produces the same analysis means as the optimal KF.
RIP accelerates the spin-up
€
xn = xn−1 +α
σn2 = Cσn−1
2
![Page 13: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/13.jpg)
RIP as mulD-‐step analysis correcDon
• With N iteration, the estimated error variance from KF-RIP is N times smaller than the one from KF.
• The estimated variance from KF-RIP is the same as the one that would be computed from KF with higher observation accuracy, i.e., with an observation error variance divided by N, the number of the iterations.
• The small estimated error variance from KF-RIP is used to achieve small increment for multi-step analysis correction
€
σa2
€
σa2 2
€
σa2 10
![Page 14: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/14.jpg)
Increase the influence of observaDons by reducing their error covariance R ☐“Hard way”:
– Reduce the observa9on error and assimilate this observa9on once.
– Compute the analysis increment at once
“soQ way”: (RIP/QOL) – Use the original observa9on error and assimilate the same observa9on mul9ple 9mes.
– The total analysis increment is achieved as the sum of mul9ple smaller increments (advantageous with nonlinear cases).
![Page 15: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/15.jpg)
Standard LETKF framework
LETKF (ti-1)
xa0 (ti-1)
xa0 (ti)
LETKF (ti)
Nonlinear model M[xa(ti-1)]
xb0 (ti)
Obs(ti)
15
Can we adjust dynamical evolu9ons at earlier 9me?
Obs(ti-1)
![Page 16: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/16.jpg)
• No-‐cost LETKF smoother ( ): apply at tn-1 the same weights found optimal at tn , works for 3D- or 4D-LETKF • Propagate information about the observation and “error of day” at tn to tn-1
★
tn-1 tn
¢
¢
¢
¢ ★
€
w a = ˜ P aYbTR−1(y −H(x ));
Wa = [(K −1) ˜ P a ]12
€
˜ x a (tn−1) = x a (tn−1) + Xa (tn−1)w a (tn )˜ X a (tn−1) = Xa (tn−1)Wa (tn )
![Page 17: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/17.jpg)
€
x a
l (tn ) = x b
l (tn ) + Xbl (tn )w (tn )
Xa
l (tn ) = Xbl (tn )W(tn )
Analysis step at tn
€
˜ x a
l (tn−1) = ˜ x a
l−1(tn−1) + ˜ X al−1(tn−1)w (tn )
˜ X a
l (tn−1) = ˜ X al−1(tn−1)W(tn ) + El−1
Smooth step at tn-1
€
xb
l (tn ) = Mtn−1 → tn[˜ x
a
l−1(tn−1) + ˜ X al−1(tn−1)]Forecast time from tn-1 to tn
threshold
€
ε =y
o(tn ) −H[x
b
l−1(tn )] − yo
l (tn ) −H[x b
l (tn )]σ o
> εs
IteraDve algorithm for re-‐using observaDons (no-‐cost smoother + stopping criterion)
The iteration continues only if we can extract extra information from the same observations
![Page 18: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/18.jpg)
“Running in place” in the LETKF framework
Xa0 (ti-1)
LETKF (ti-1)
xa(ti) False
xbn (ti)
�
� �
Random Pert. +
Obs(ti)
�
�
5/14
Time
Obs(ti-1)
Nonlinear model M [xa(ti-1) ]
LETKF (ti) Threshold > ε
no-cost Smoother
€
˜ x an (ti−1)
re-‐evolve the whole ensemble to catch up the true dynamics, represented by OBS
![Page 19: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/19.jpg)
“Quasi Outer-‐loop” in the LETKF framework (simplified RIP)
Xa0 (ti-1)
LETKF (ti-1)
no-cost Smoother
False
LETKF (ti) Threshold > ε
Nonlinear model M [ xa(ti-1) ] �
�
� Obs(ti)
�
�
5/14
Time
Obs(ti-1)
€
x bn (ti)
€
x an (ti)
€
x an (ti−1)
adjust the nonlinearity of the mean trajectory
Random Pert. +
![Page 20: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/20.jpg)
We’ll focus on nonlinear dynamics, and propose two new methods based on the LETKF framework for using long windows
Dealing with nonlineari9es with EnKF
RIP QOL generalized outer-‐loop
simplified RIP
improvement mean and covariance mean
cost expensive less expensive
itera9on number < 10 1~2
![Page 21: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/21.jpg)
Experimental sepng
• Nonlinear model: Lorenz 3-variable model • assimilation setup
– DA methods: LETKF, RIP, QOL and 4D-Var – observation error variance= 2.0 – assimilation window
• Linear window (frequent observation): 8 timestep • Nonlinear window (infrequent observation): 25 timestep
![Page 22: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/22.jpg)
Results with Lorenz 3-‐variable model
4D-‐Var LETKF
Linear window (obs every 8 Dmesteps)
0.31 0.30
Nonlinear window (obs every 25 Dmesteps)
0.53 Assim
window=75 0.68
• Long window + Quasi-‐sta9c varia9onal analysis (Pires et al.,1996) -‐> 4D-‐Var wins! • The standard LETKF can’t handle the long assim. window.
![Page 23: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/23.jpg)
Results with Lorenz 3-‐variable model
4D-Var LETKF
standard +QOL +RIP obs every 8 time-‐step (linear window) 0.31 0.30 0.27 0.27
obs every 25 time-‐step (nonlinear window)
0.53 (assim window=75) 0.68 0.47 0.35
• With the QOL, LETKF analysis with nonlinear window is much improved, even bexer than 4D-‐Var!
• RIP gives even more improvement than the QOL because it improves both the mean and the covariance.
23
![Page 24: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/24.jpg)
Trajectory of variable y
with RIP/QOL filter divergence is avoided
![Page 25: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/25.jpg)
fewer observa9ons x y z xy xz yz xyz
ETKF 2.9 1.67 7.16 1.01 1.53 0.78 0.68
QOL 1.98 1.23 5.94 0.82 1.16 0.60 0.47
RIP 1.57 0.97 3.81 0.56 0.66 0.40 0.35
• With fewer observa9ons (constraint), the model trajectory is strongly affected by the nonlinear evolu9on of the ini9al errors. • RIP and QOL use the observa9ons more efficiently for the under-‐observed cases. • Performance: RIP > QOL > standard ETKF
![Page 26: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/26.jpg)
Comparisons between itera9ve EnKFs
1. RIP/QOL 2. Ensemble Randomized Maximum Likelihood
(EnRML, Gu and Oliver, 2007) – Same framework as the 4D-‐Var: Improve only the sensi9vity matrix, Hx, and the background trajectory for compu9ng the innova9on, yo-‐H(xb)
– Minimize the cost-‐func9on with the reduced adjustment Gauss-‐Newton method
![Page 27: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/27.jpg)
Ensemble Randomized Maximum Likelihood Implement MLH with the stochas9c EnKF
(1) Minimizing the cost-‐func9on is solved for the ensemble member (k), with perturbed observa9ons at tn
(2) es9mate the data mismatch for both and with OMFi
€
x0k,i+1 x0
k,i
(3) if OMFl+1< OMFl, and increase βl, otherwise keep and decrease βl
€
x0k,i+1 = x0
k,i
€
x0k,l
(4) If the criteria is not sa9sfied, repeat (1) to (3). Criteria to stop the itera9on:
• (OMFl − OMFl+1)/OMFl < 10-‐4 • Maximum itera9on number is 20
€
J(x0k ) =
12x0k − xb0
k[ ]TPb0−1 x0k − x k0k[ ]+12H (xn
k )− yonk[ ]TR −1 H (xn
k )− yonk[ ]
x0k, i+1 = βixb0
k + (1− βi )x0ki − βi
1K −1⎛
⎝ ⎜
⎞
⎠ ⎟ X b0
k (Ybnk )T(R +
1K −1⎛
⎝ ⎜
⎞
⎠ ⎟ Ybn
k (Ybnk )T xn
k, i − yonk, i − (xn
k, i − xbnk )[ ]
€
OMFi =12H(M[x0
k,i]) − yonk[ ]
TR−1 H(M[x0
k,i]) − yonk[ ]
![Page 28: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/28.jpg)
RMS error with itera9ve EnKFs (K=24)
QOL RIP EnRML
W/O adjus9ng the minimizing step
RMS error 0.49 0.33 0.41 1.45
• Experiment with EnRML is performed with an assimila9on window of 25 9me-‐step with observa9ons arranged at the end of the window • With infrequent observa9on, RIP performs bexer than EnRML.
(EnRML doesn’t work with 3 ensemble member, while RIP and QOL already reach opDmal value at K=3)
![Page 29: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/29.jpg)
• The minimiza9on of EnRML can s9ll fail with strong nonlinearity • RIP/QOL follow the true trajectory bexer • However, when EnRML is well behaved, it is slightly more accurate than RIP/QOL.
![Page 30: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/30.jpg)
Applica9on of LETKF-‐RIP to typhoon assimila9on/predic9on
06! 09! 12!00! 03! 15! 18!time!
A B
LETKF-RIP setup!1) Computed the LETKF weights at analysis time (00,06,12,18Z)"2) Use these weight to reconstruct the ensemble (U, V) at (03,09,15,21Z)"3) perform the 3-hr ensemble forecasts"4) Re-do the LETKF analysis (only one iteration is tested)"
OSSE experiment setup: • Regional Model: Weather Research and Forecasting model (WRF, 25km) • Assimilation scheme: LETKF and LETKF-RIP with 36 ensemble members • Observations: radiosonde, dropsondes and surface ocean wind
![Page 31: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/31.jpg)
Typhoon ver9cal structure
(analysis)
TRUTH LETKF LETKF-RIP
Time
Rapidly intensifying
Color: Wind speed
![Page 32: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/32.jpg)
Typhoon predic9on: represent the environmental condi9on
Truth LETKF-‐RIP LETKF
• LETKF-RIP is able to accelerate the adjustment of the environmental condition for typhoon development: • When initialized with the LETKF-RIP analysis, the improvements include:
1. Capture the west-ward turning direction the typhoon track.
2. Capture the slow typhoon movement speed when approaching Taiwan
Translation speed of typhoon
![Page 33: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/33.jpg)
Typhoon predic9on: typhoon intensity
• The typhoon intensity can be also spun-up by the LETKF-RIP • The advantage is still valid for the 24-hour forecast
![Page 34: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/34.jpg)
Steve Penny’s defense April 15, 2011
l Eugenia Kalnay
l Jim Carton
l Brian Hunt
l Kayo Ide
l Takemasa Miyoshi
l Gennady Chepurin
With:
An application of LETKF-RIP to ocean data assimilation
![Page 35: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/35.jpg)
0.00
0.50
1.00
1.50
2.00
2.50
Jan-‐97 Jan-‐98 Jan-‐99 Jan-‐00 Jan-‐01 Jan-‐02 Jan-‐03 Jan-‐04
LETKF-RIP A
Free-‐Run
LETKF-IAU B SODA B SODA A LETKF-IAU A
LETKF-RIP B
RMSD (ºC) (All vertical levels)
Ocean Reanalysis (7 years): LETKF-IAU, LETKF-RIP, compared with SODA (OI)
12-month running mean
Temperature RMS differences wrt obs
![Page 36: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/36.jpg)
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
Jan-‐97 Jan-‐98 Jan-‐99 Jan-‐00 Jan-‐01 Jan-‐02 Jan-‐03 Jan-‐04
LETKF-RIP B./A.
Free-‐Run
LETKF-IAU B. SODA B. / A.
LETKF-IAU A.
RMSD (psu) (All vertical levels)
Salinity RMS differences wrt obs
Ocean Reanalysis (7 years): LETKF-IAU, LETKF-RIP, compared with SODA (OI)
12-month running mean
![Page 37: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/37.jpg)
Summary • As in the varia9onal methods, an outer-‐loop with LETKF
(EnKF) allows to improve the nonlinear evolu9on of the background trajectory and bexer fit the observa9ons. – Both the RIP and QOL methods are able to improve the nonlinearity of
the model trajectory, so less non-‐Gaussian distribu9on occurs.
• Despite viola9ng the Kalman Filter rule that observa9ons should be used only once, the QOL and RIP methods are clearly very successful to use observa9on more than once. – Mul9-‐step analysis correc9on with small ensemble spread
• The RIP analysis is actually more accurate than the EnRML analysis, a itera9ve EnKF based on Gauss-‐Newton minimiza9on.
![Page 38: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/38.jpg)
Running in place (Kalnay and Yang, 2011)
• During the spin-‐up, we propose to use the observations repeatedly “ONLY IF” we could extract extra information. But we should avoid overXitting the observations.
• With RIP, we improve both the accuracy of the mean state and the Xlow-‐dependent error structures.
• Elements for RIP – No-‐cost smoother (vs. adjoint model in 4D-‐Var) – An appropriate scheme to avoid over-‐Xitting
38
![Page 39: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/39.jpg)
RIP-‐LETKF with the QG model (Kalnay and Yang, 2010)
39
RMS error
Analysis error of potential vorticity of a QG model
![Page 40: Ensemble(Kalman(Filter:( Handlingnonlinearitywiththe ......Jun 13, 2011 · Smooth step at t n-1 € x b l(t n)=M t n−1→t n [x ˜ a l−1(t n−1)+ X ˜ a l−1(t n−1)] Forecast](https://reader036.fdocuments.net/reader036/viewer/2022071504/6123ee6006a765467013e13d/html5/thumbnails/40.jpg)
RIP-‐LETKF with the QG model (Kalnay and Yang, 2009)
40
RMS error
Analysis error of potential vorticity of a QG model
• LETKF spin-up from random perturbations: 141 cycles. With RIP: 46 cycles • LETKF spin-up from 3D-Var perts 54 cycles. With RIP: 37 cycles • 4D-Var spin-up using 3D-Var prior: 54 cycles