introduction to data assimilation - cirm · introduction to data assimilation sebastian reich ()...
TRANSCRIPT
Introduction to data assimilation
Sebastian Reich (www.sfb1294.de)
Universität Potsdam/ University of Reading
CIRM2018, October 26, 2018
Universität Potsdam/ University of Reading 1
Ensemble prediction I
Source: The quiet revolution of numerical weather prediction, Nature, 2015
Universität Potsdam/ University of Reading 2
Ensemble prediction II
Ensemble prediction system with M members:d
dtZit
= f (Zit) , Zi0 ∼ π0 , i = 1, . . . ,M .
Source: ECMWF
Universität Potsdam/ University of Reading 3
Data assimilation I
Source: The quiet revolution of numerical weather prediction, Nature, 2015
Universität Potsdam/ University of Reading 4
Data assimilation II
Source: The quiet revolution of numerical weather prediction, Nature, 2015
Universität Potsdam/ University of Reading 5
Data assimilation III
StateState StateModel Model
Data
Assimilation
Data
Assimilation
Data
Assimilation
time
Model:dZt = f (Zt)dt + γ1/2dWt
Data: ytn at discrete times tn
Universität Potsdam/ University of Reading 6
Data assimilation IV
Discrete–time observations:
ytn = h(Ztn) +R1/2 Ξtn , n = 1, . . . ,N .
Likelihood function:
L(z[0,T]) := exp
�
−1
2
∑
n
‖ytn − h(ztn)‖2R
�
.
Bayes:dbP[0,T]
dP[0,T](Z[0,T]) :=
L(Z[0,T])
P[0,T][L].
Universität Potsdam/ University of Reading 7
Data assimilation V
prior P[9:21] in yellow; posterior bP[9:21] in green
Universität Potsdam/ University of Reading 8
References
É Daley, R., Atmospheric Data Analysis, Cambridge University Press,1991.
É Bain, A. and Crisan, D., Fundamentals of Stochastic Filtering,Springer, 2009.
É SR, Cotter, J., Probabilistic Forecasting and Bayesian DataAssimilation, Cambridge University Press, 2015.
É Law, K. et al., Data Assimilation – A Mathematical Introduction,Springer, 2015.
É P. Bauer et al, The quiet revolution of numerical weather prediction,Nature, 2015
Universität Potsdam/ University of Reading 9
Sequential processing of data I
StateState StateModel Model
Data
Assimilation
Data
Assimilation
Data
Assimilation
time
Notation:
Prediction/Forecast: π(bzn+1 |y1:n)
Filtering/Analysis: π(zn+1 |y1:n+1)
Universität Potsdam/ University of Reading 10
Sequential processing of data II
Classic bootstrap particle filter (sequential MC):
1) propagate M particles zin
under the model dynamics to yield forecastsbzin+1, i = 1, . . . ,M.
2) data likelihood assigns importance weights win+1 to each
propagated particle bzin+1
3) resample to obtain equally weighted particles zin+1 from the filtering
distribution at tn+1.
This approach is not applicable to geophysical/meteorologicalapplications because of:
i) small sample sizes M ∼ 102
ii) high dimensionality of models (discretised PDEs) D ∼ 107
iii) medium number of observations N ∼ 105
Universität Potsdam/ University of Reading 11
Sequential processing of data II
Classic bootstrap particle filter (sequential MC):
1) propagate M particles zin
under the model dynamics to yield forecastsbzin+1, i = 1, . . . ,M.
2) data likelihood assigns importance weights win+1 to each
propagated particle bzin+1
3) resample to obtain equally weighted particles zin+1 from the filtering
distribution at tn+1.
This approach is not applicable to geophysical/meteorologicalapplications because of:
i) small sample sizes M ∼ 102
ii) high dimensionality of models (discretised PDEs) D ∼ 107
iii) medium number of observations N ∼ 105
Universität Potsdam/ University of Reading 11
Sequential processing of data III
Remedies used in data assimilation:
1) Formulate assimilation of data as an optimisation problem (4DVar,3DVar)
2) Replace importance–resampling step by low variance ensembletransformations
zjn+1 =
M∑
i=1
bzin+1 t
∗ij
3) Localise assimilation of data
zjn+1(x) =
M∑
i=1
zin+1(x) t∗
ij(x)
x ∈ R3 grid point beingupdated.
State variable on grid points
Observations outside localization radius
Gridpoint being updated
Observations inside localization radius
Universität Potsdam/ University of Reading 12
Sequential processing of data III
Remedies used in data assimilation:
1) Formulate assimilation of data as an optimisation problem (4DVar,3DVar)
2) Replace importance–resampling step by low variance ensembletransformations
zjn+1 =
M∑
i=1
bzin+1 t
∗ij
3) Localise assimilation of data
zjn+1(x) =
M∑
i=1
zin+1(x) t∗
ij(x)
x ∈ R3 grid point beingupdated.
State variable on grid points
Observations outside localization radius
Gridpoint being updated
Observations inside localization radius
Universität Potsdam/ University of Reading 12
Sequential processing of data III
Remedies used in data assimilation:
1) Formulate assimilation of data as an optimisation problem (4DVar,3DVar)
2) Replace importance–resampling step by low variance ensembletransformations
zjn+1 =
M∑
i=1
bzin+1 t
∗ij
3) Localise assimilation of data
zjn+1(x) =
M∑
i=1
zin+1(x) t∗
ij(x)
x ∈ R3 grid point beingupdated.
State variable on grid points
Observations outside localization radius
Gridpoint being updated
Observations inside localization radius
Universität Potsdam/ University of Reading 12
Ensemble Kalman filter (EnKF)
Forward observation model: yobsn+1 = h(bzn+1) +R1/2Ξn+1
Best linear unbiased estimator (BLUE): zn+1 = Ayobsn+1 + b
Empirical BLUE/EnKF:
A = PR−1
bj = bzjn+1 − A
�
h(bzjn+1) +R1/2
bΞjn+1
�
with
P =1
M− 1
M∑
i=1
bzin+1(h(bzi
n+1)− hn+1)T , bΞin+1 ∼ N(0, I) .
EnKF induced particle transformation:
zjn+1 =
M∑
i=1
bzin+1t
∗ij.
Universität Potsdam/ University of Reading 13
Ensemble Kalman filter (EnKF)
Forward observation model: yobsn+1 = h(bzn+1) +R1/2Ξn+1
Best linear unbiased estimator (BLUE): zn+1 = Ayobsn+1 + b
Empirical BLUE/EnKF:
A = PR−1
bj = bzjn+1 − A
�
h(bzjn+1) +R1/2
bΞjn+1
�
with
P =1
M− 1
M∑
i=1
bzin+1(h(bzi
n+1)− hn+1)T , bΞin+1 ∼ N(0, I) .
EnKF induced particle transformation:
zjn+1 =
M∑
i=1
bzin+1t
∗ij.
Universität Potsdam/ University of Reading 13
Ensemble Kalman filter (EnKF)
Forward observation model: yobsn+1 = h(bzn+1) +R1/2Ξn+1
Best linear unbiased estimator (BLUE): zn+1 = Ayobsn+1 + b
Empirical BLUE/EnKF:
A = PR−1
bj = bzjn+1 − A
�
h(bzjn+1) +R1/2
bΞjn+1
�
with
P =1
M− 1
M∑
i=1
bzin+1(h(bzi
n+1)− hn+1)T , bΞin+1 ∼ N(0, I) .
EnKF induced particle transformation:
zjn+1 =
M∑
i=1
bzin+1t
∗ij.
Universität Potsdam/ University of Reading 13
Ensemble transform particle filter
Importance weights: win+1 ∝ exp
�
−1
2‖h(bzi
n+1)− yn+1)‖2R
�
5 10 15 200
0.02
0.04
0.06
0.08
0.1
Impo
rta
nce
we
igth
Particle Number
⇒OptimalCoupling
T∗⇒
5 10 15 200
0.02
0.04
0.06
0.08
0.1
Impo
rta
nce
we
igth
Particle Number
T∗ minimiser of J({tij}) =M∑
i,j=1
tij‖bzin+1 − bzjn+1‖
2
subject to:tij ≥ 0,
∑
i
tij = 1,∑
j
tij = wiM
T∗ induces linear transformation: zjn+1 =
M∑
i=1
bzin+1t
∗ij
Universität Potsdam/ University of Reading 14
Ensemble transform particle filter
Importance weights: win+1 ∝ exp
�
−1
2‖h(bzi
n+1)− yn+1)‖2R
�
5 10 15 200
0.02
0.04
0.06
0.08
0.1
Imp
ort
an
ce
we
igth
Particle Number
⇒OptimalCoupling
T∗⇒
5 10 15 200
0.02
0.04
0.06
0.08
0.1
Impo
rta
nce
we
igth
Particle Number
T∗ minimiser of J({tij}) =M∑
i,j=1
tij‖bzin+1 − bzjn+1‖
2
subject to:tij ≥ 0,
∑
i
tij = 1,∑
j
tij = wiM
T∗ induces linear transformation: zjn+1 =
M∑
i=1
bzin+1t
∗ij
Universität Potsdam/ University of Reading 14
Ensemble transform particle filter
Importance weights: win+1 ∝ exp
�
−1
2‖h(bzi
n+1)− yn+1)‖2R
�
5 10 15 200
0.02
0.04
0.06
0.08
0.1
Imp
ort
an
ce
we
igth
Particle Number
⇒OptimalCoupling
T∗⇒
5 10 15 200
0.02
0.04
0.06
0.08
0.1
Imp
ort
an
ce
we
igth
Particle Number
T∗ minimiser of J({tij}) =M∑
i,j=1
tij‖bzin+1 − bzjn+1‖
2
subject to:tij ≥ 0,
∑
i
tij = 1,∑
j
tij = wiM
T∗ induces linear transformation: zjn+1 =
M∑
i=1
bzin+1t
∗ij
Universität Potsdam/ University of Reading 14
Ensemble transform particle filter
Importance weights: win+1 ∝ exp
�
−1
2‖h(bzi
n+1)− yn+1)‖2R
�
5 10 15 200
0.02
0.04
0.06
0.08
0.1
Imp
ort
an
ce
we
igth
Particle Number
⇒OptimalCoupling
T∗⇒
5 10 15 200
0.02
0.04
0.06
0.08
0.1
Imp
ort
an
ce
we
igth
Particle Number
T∗ minimiser of J({tij}) =M∑
i,j=1
tij‖bzin+1 − bzjn+1‖
2
subject to:tij ≥ 0,
∑
i
tij = 1,∑
j
tij = wiM
T∗ induces linear transformation: zjn+1 =
M∑
i=1
bzin+1t
∗ij
Universität Potsdam/ University of Reading 14
Ensemble transform particle filter
Importance weights: win+1 ∝ exp
�
−1
2‖h(bzi
n+1)− yn+1)‖2R
�
5 10 15 200
0.02
0.04
0.06
0.08
0.1
Imp
ort
an
ce
we
igth
Particle Number
⇒OptimalCoupling
T∗⇒
5 10 15 200
0.02
0.04
0.06
0.08
0.1
Imp
ort
an
ce
we
igth
Particle Number
T∗ minimiser of J({tij}) =M∑
i,j=1
tij‖bzin+1 − bzjn+1‖
2
subject to:tij ≥ 0,
∑
i
tij = 1,∑
j
tij = wiM
T∗ induces linear transformation: zjn+1 =
M∑
i=1
bzin+1t
∗ij
Universität Potsdam/ University of Reading 14
Ensemble transform particle filter
Importance weights: win+1 ∝ exp
�
−1
2‖h(bzi
n+1)− yn+1)‖2R
�
5 10 15 200
0.02
0.04
0.06
0.08
0.1
Imp
ort
an
ce
we
igth
Particle Number
⇒OptimalCoupling
T∗⇒
5 10 15 200
0.02
0.04
0.06
0.08
0.1
Imp
ort
an
ce
we
igth
Particle Number
T∗ minimiser of J({tij}) =M∑
i,j=1
tij‖bzin+1 − bzjn+1‖
2
subject to:tij ≥ 0,
∑
i
tij = 1,∑
j
tij = wiM
T∗ induces linear transformation: zjn+1 =
M∑
i=1
bzin+1t
∗ij
Universität Potsdam/ University of Reading 14
Numerical example I
Lorenz-63 model, first component observed infrequently (∆t = 0.12) andwith large measurement noise (R = 8):
Figure: RMSEs for various second-order accurate LETFs compared to the ETPF, theESRF, and the SIR PF as a function of the sample size, M.
Universität Potsdam/ University of Reading 15
Numerical example II
Hybrid filter: P := PESRF(α) PETPF(1− α).
Figure: RMSEs for hybrid ESRF (α = 0) and 2nd-order corrected NETF/ETPF (α = 1)as a function of the sample size, M.
Universität Potsdam/ University of Reading 16
References
É Evensen, G., Data assimilation – the ensemble Kalman filter,Springer, 2009.
É SR, A nonparametric ensemble transform method for Bayesianinference, SIAM J. Sci. Comput., 2013.
É SR, Cotter, J., Probabilistic Forecasting and Bayesian DataAssimilation, Cambridge University Press, 2015.
É Tödter J., Ahrens, B., A second–order exact ensemble square rootfilter for nonlinear data assimilation, Month. Weather Rev., 2015.
É Chustagulprom, N. et al., A hybrid ensemble transform particle filterfor nonlinear and spatially extended systems, SIAM/ASA J. UQ, 2016.
É Acevedo, W., et al., Second–order accurate ensemble transformparticle filters, SIAM J. Sci. Comput., 2017.
É Fearnhead, P., Künsch, H.R., Particle filters and data assimilation,Annual Review of Statistics and Its Application, 2018.
Universität Potsdam/ University of Reading 17
Continuous–time data assimilation I
Observation model:
dyt = h(Zt)dt +R1/2dVt
Linear SDEs and observations:
dZt = (AZt + b) dt + γ1/2dWt, dyt = HZtdt +R1/2dVt .
Kalman-Bucy equations for the mean
dμt = (Aμt + b)dt − KtdIt, dIt := (Hμtdt − dyt)
and the covariance matrix
d
dtPt = APt + PtA
T + γI− KtHPt .
with Kalman gain matrix Kt = PtHTR−1.
Universität Potsdam/ University of Reading 18
Continuous–time data assimilation I
Observation model:
dyt = h(Zt)dt +R1/2dVt
Linear SDEs and observations:
dZt = (AZt + b) dt + γ1/2dWt, dyt = HZtdt +R1/2dVt .
Kalman-Bucy equations for the mean
dμt = (Aμt + b)dt − KtdIt, dIt := (Hμtdt − dyt)
and the covariance matrix
d
dtPt = APt + PtA
T + γI− KtHPt .
with Kalman gain matrix Kt = PtHTR−1.
Universität Potsdam/ University of Reading 18
Continuous–time data assimilation I
Observation model:
dyt = h(Zt)dt +R1/2dVt
Linear SDEs and observations:
dZt = (AZt + b) dt + γ1/2dWt, dyt = HZtdt +R1/2dVt .
Kalman-Bucy equations for the mean
dμt = (Aμt + b)dt − KtdIt, dIt := (Hμtdt − dyt)
and the covariance matrix
d
dtPt = APt + PtA
T + γI− KtHPt .
with Kalman gain matrix Kt = PtHTR−1.
Universität Potsdam/ University of Reading 18
Continuous–time data assimilation II
Controlled mean–field SDE:
dZt = (AZt + b)dt + γ1/2dWt + dut(Zt), dut := −KtdIt
with innovation It
dIt =1
2(HZt +Hμt)dt − dyt
ordIt = HZtdt +R1/2dUt − dyt
and Kalman gain Kt = PtHTR−1.
Extension to nonlinear SDEs:
Ensemble Kalman-Bucy filter(EnKBF)
dtZt = f (Zt)dt + γ1/2dWt − KtdIt .
Universität Potsdam/ University of Reading 19
Continuous–time data assimilation II
Controlled mean–field SDE:
dZt = (AZt + b)dt + γ1/2dWt + dut(Zt), dut := −KtdIt
with innovation It
dIt =1
2(HZt +Hμt)dt − dyt
ordIt = HZtdt +R1/2dUt − dyt
and Kalman gain Kt = PtHTR−1.
Extension to nonlinear SDEs:
Ensemble Kalman-Bucy filter(EnKBF)
dtZt = f (Zt)dt + γ1/2dWt − KtdIt .
Universität Potsdam/ University of Reading 19
Continuous–time data assimilation III
Feedback particle filter:
dZt = f (Zt)dt + γ1/2dWt − Kt ◦ dIt
with Kalman gain
Kt := ∇zψtR−1, ∇z · (πt∇zψt) = −πt(h− h)
and innovations
dIt =1
2(h(Zt) + ht)dt − dyt
ordIt = h(Zt)dt +R1/2dUt − dyt .
Universität Potsdam/ University of Reading 20
Continuous–time data assimilation IV
Particle approximation to elliptic PDE:
∇zψt(zjt) ≈
M∑
i=1
zitaijt
with zit∼ πt, i = 1, . . . ,M and coefficients a
ijt ∈ R appropriately defined.
Interacting particle system:
dzjt = f (zjt)dt + γ1/2dWj
t −
M∑
i=1
zitaijt
!
R−1 ◦ dI jt
with innovationdI jt = h(z
jt)dt +R1/2dUjt − dyt .
Universität Potsdam/ University of Reading 21
Continuous–time data assimilation IV
Particle approximation to elliptic PDE:
∇zψt(zjt) ≈
M∑
i=1
zitaijt
with zit∼ πt, i = 1, . . . ,M and coefficients a
ijt ∈ R appropriately defined.
Interacting particle system:
dzjt = f (zjt)dt + γ1/2dWj
t −
M∑
i=1
zitaijt
!
R−1 ◦ dI jt
with innovationdI jt = h(z
jt)dt +R1/2dUjt − dyt .
Universität Potsdam/ University of Reading 21
Parameter estimation I
Parameter–dependent model:
dZt = f (Zt, θ)dt + γ1/2dWt − Kt ◦ dIt
Marginal likelihood (evidence of the data Yt = y[0,t] under θ):
dπt(Yt |θ) = πt(Yt |θ) htR−1dyt .
Example.SPDE model:
ut = θ∆ut + Wt
Wt space–time white noise.
Observation:dyt = ut(x0) dt +R1/2dVt
for fixed spatial location x0.
Universität Potsdam/ University of Reading 22
Parameter estimation I
Parameter–dependent model:
dZt = f (Zt, θ)dt + γ1/2dWt − Kt ◦ dIt
Marginal likelihood (evidence of the data Yt = y[0,t] under θ):
dπt(Yt |θ) = πt(Yt |θ) htR−1dyt .
Example.SPDE model:
ut = θ∆ut + Wt
Wt space–time white noise.
Observation:dyt = ut(x0) dt +R1/2dVt
for fixed spatial location x0.
Universität Potsdam/ University of Reading 22
Parameter estimation II
D = 80 grid points, state estimation with EnKBF, M = 20 particles, andlocalisation, observation interval [0,20], measurement error R = 0.001
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8parameter
-32
-30
-28
-26
-24
-22
-20
-18
-16
-14
-12ne
gativ
e lo
glik
elih
ood
Figure: negative loglikelihood for θ ∈ [0.2,1.8], true value θ∗ = 1.Universität Potsdam/ University of Reading 23
References
É Bergemann, K, SR, An ensemble Kalman–Bucy filter forcontinuous–time data assimilation, Meteorolog. Zeitschrift, 2012.
É Yang, T. et al., Feedback particle filter, IEEE Trans. Autom. Control,2013.
É Taghvaei, A. et al., Kalman filter and its modern extensions for thecontinuous–time nonlinear filtering problem, J. Dyn. Sys., Meas., andControl, 2018.
É de Wiljes, et al., Long–time stability and accuracy of the ensembleKalman–Bucy filter for fully observed processes and smallmeasurement noise, SIAM J. Appl. Dyn. Sys., 2018.
É SR, Data assimilation, Acta Numerica, 2019, preliminary version onarXiv 1807.08351.
Universität Potsdam/ University of Reading 24
Summary
É Data assimilation has become a workhorse for day–to–day weatherforecasting
É Key are a "smart" choice of biased and robust approximations to thesequential filtering problem
É Data assimilation has applications to many areas outside weatherprediction such as medicine, geosciences, biology, etc.
É Current topics of research: combined state–parameter estimation forPDEs, theoretical understanding of approximations such as lineartransformations & localisation
É Continuous–in–time DA naturally leads to an attractive unifyingframework in the form of mean–field SDEs.
Universität Potsdam/ University of Reading 25
The End
(source: J.B. Handelsman, New Yorker, 05/31/1969)
Universität Potsdam/ University of Reading 26
Collaborators
É Walter Acevedo
É Kay Bergemann
É Yuan Cheng
É Nawinda Chustagulprom
É Colin Cotter
É Jana de Wiljes
É Prashant Mehta
É Wilhelm Stannat
É Amari Taghvaei
Universität Potsdam/ University of Reading 27