the combined parameter and state estimation problem · the combined parameter and state estimation...
TRANSCRIPT
![Page 1: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/1.jpg)
The combined parameter and state
estimation problemGeir Evensen and Laurent Bertino
Hydro Research Centre, Bergen, Norway,Nansen Environmental and Remote Sensing Center, Bergen, Norway
![Page 2: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/2.jpg)
Combined parameter and state estimation
“Find the pdf of the parameters and the model solutionconditional on a set of measurements.”
Bayesian formulation.
Ensemble solutions.
![Page 3: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/3.jpg)
Bayesian formulation
Model equations and measurements:
∂ψ
∂t= g(ψ,α) + q,
ψ|t0 = Ψ0 + a,
α = α0 +α′,
M(ψ,α) = d+ ε.
Bayes theorem becomes:
f(ψ,α,ψ0|d) ∝ f(ψ|α,ψ0)f(ψ0)f(α) f(d|ψ,α).
![Page 4: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/4.jpg)
Gaussian priors
f(ψ,α,ψ0|d) ∝ exp
{
−1
2J [ψ,α]
}
,
J [ψ,α] =
(
∂ψ
∂t− g(ψ,α)
)T
•W qq •
(
∂ψ
∂t− g(ψ,α)
)
+ (ψ0 − Ψ0)T ◦W aa ◦ (ψ0 − Ψ0)
+ (α−α0)T ◦W αα ◦ (α−α0)
+ (d− M[ψ,α])TW εε(d− M[ψ,α]).
MLH solution, hard to solve and to compute error estimates.
![Page 5: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/5.jpg)
Discretisation in time
.
-
t0 t1 t2 t3 t4 t5 t6 t7 tk
ψ1 ψ2 ψ3 ψ4 ψ5 ψ6 ψ7 ψkψ0
dmd2 d3d1
ti(1) ti(2) ti(m)
djti(j)
ti
ψi
ti(3)
. . .
. . .
. . .. . .
. . . . . .
. . .
. . .
![Page 6: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/6.jpg)
Assume
Model is first order Markov process.
f(ψ1, . . . ,ψk,α,ψ0) ∝
f(α)f(ψ0)
k∏
i=1
f(ψi|ψi−1,α).
Measurement errors are independent in time
f(d|ψ,α) =
m∏
j=1
f(dj |ψi(j),α).
![Page 7: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/7.jpg)
Bayes for discrete state
Bayes theorem then becomes
f(ψ1, . . . ,ψk,α,ψ0|d) ∝ f(α)f(ψ0)
k∏
i=1
f(ψi|ψi−1,α)
m∏
j=1
f(dj |ψi(j),α),
Time
d_2d_1 d_m
Prediction
Update
![Page 8: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/8.jpg)
Rewrite as:
f(ψ1, . . . ,ψk,α,ψ0|d) ∝ f(α)f(ψ0)
i(1)∏
i=1
f(ψi|ψi−1,α)f(d1|ψi(1),α)
i(2)∏
i=i(1)+1
f(ψi|ψi−1,α)f(d2|ψi(2),α) · · ·
i(m)∏
i=i(m−1)+1
f(ψi|ψi−1,α)f(dm|ψi(m),α)
k∏
i=i(m)+1
f(ψi|ψi−1,α)
![Page 9: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/9.jpg)
Sequential processing of measurements
First update
f(ψ1, . . . ,ψi(1),α,ψ0|d1) ∝
f(α)f(ψ0)
i(1)∏
i=1
f(ψi|ψi−1,α)f(d1|ψi(1),α)
Second update
f(ψ1, . . . ,ψi(2),α,ψ0|d1,d2) ∝
f(ψ1, . . . ,ψi(1),α,ψ0|d1)
i(2)∏
i=i(1)+1
f(ψi|ψi−1,α)f(d2|ψi(2),α)
![Page 10: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/10.jpg)
Summary
Independent measurements processed sequentially in time.
Sequence of inverse problems.
Solution of one sub-problem is prior for next.
Hard to solve using traditional variational methods.
Well suited for ensemble methods.
Time
Updates
d_3d_1 d_2
![Page 11: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/11.jpg)
Variance minimizing analysis
Given
First guess, ψf , and prediction error covariance, C fψψ.
Measurements, d, and error covariance, Cεε.
The variance minimizing analysis is minimum of
J [ψa] =(
ψa −ψf)T
•C−1ψψ •
(
ψa −ψf)
+(
d− Mψa)TC−1εε
(
d− Mψa)
.
![Page 12: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/12.jpg)
Analysis equations
Update of estimate
ψa = ψf +CfψψM
T(
MCfψψM
T +Cεε
)
−1(
d− Mψf)
Error covariance for update
Caψψ = Cf
ψψ −CfψψM
T(
MCfψψM
T +Cεε
)
−1MC f
ψψ
![Page 13: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/13.jpg)
Ensemble methods
ES: Ensemble Smoother
EnKS: Ensemble Kalman Smoother
EnKF: Ensemble Kalman Filter
Ensemble representation for pdfs.
Ensemble prediction for time evolution of pdfs.
Linear ensemble analysis scheme:“Variance minimizing”.Assumes Gaussian pdf for model prediction.No resampling!
![Page 14: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/14.jpg)
The error covariance matrix
Define ensemble covariances around the ensemble mean
Cψψ =(
ψ −ψ) (
ψ −ψ)T
The ensemble mean, ψ, is the best-guess.
The error variance is defined by the ensemble spread.
The smoothness of members defines the covariance.
⇓Representation by a finite ensemble of model states.
![Page 15: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/15.jpg)
Ensemble representation
Define ensemble matrix
A(x, ti) =
(
ψ1(x, ti) ψ2(x, ti) . . . ψN (x, ti)
α1(x, ti) α2(x, ti) . . . αN (x, ti)
)
.
Parameters augmented to model state.
The ensemble covariance becomes
Ceψψ(x1,x2, ti) =
A′(x1, ti)A′Ti (x2, ti)
N − 1.
![Page 16: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/16.jpg)
Dynamical evolution of error statistics
Members evolve according to stochastic model dynamics
dψ = g(ψ,α)dt+ h(ψ)dq
dα = 0.
The pdf evolves according to Kolmogorov’s equation
∂f
∂t+∑
i
∂(gif)
∂ψi=
1
2
∑
i,j
∂2f(
hQhT)
ij
∂ψi∂ψj.
Fundamental equation for evolution of error statistics.
May be solved using Monte Carlo methods.
![Page 17: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/17.jpg)
Ensemble analysis equations
Define randomized measurements D = d+E.
Variance minimizing ensemble update becomes
Aa = A
+CeψψM
T(
MCeψψM
T +Cεε
)
−1(
D − MA)
= A
+A′(
MA′)T(
MA′(
MA′)T
+Cεε
)
−1(
D − MA)
...
= AX
⇒ Updated ensemble with correct mean and covariance!
![Page 18: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/18.jpg)
ES: The Ensemble Smoother
Computes the Bayesian update using linear ensemble equation
f(ψ1, . . . ,ψk,α,ψ0|d) =
f(ψ1, . . . ,ψk,α,ψ0)
m∏
j=1
f(dj |ψi(j),α),
Time
d_2d_1 d_m
Prediction
Update
![Page 19: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/19.jpg)
ES: Example with Lorenz equations
-20
-15
-10
-5
0
5
10
15
20
0 5 10 15 20 25 30 35 40
ES: Estimate
0
2
4
6
8
10
0 5 10 15 20 25 30 35 40
ES: Errors
![Page 20: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/20.jpg)
ES: summary
Gauss–Markov interpolation in space and time.
Creates an ensemble for the model evolution.
Assumes Gaussian pdf for model evolution.
Computes variance minimizing ensemble analysis.
Exact solution for linear problems.
![Page 21: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/21.jpg)
EnKS: The ensemble Kalman smoother
Bayes with sequential processing of data: First update is
f(ψ1, . . . ,ψi(1),α,ψ0|d1) ∝
f(α)f(ψ0)
i(1)∏
i=1
f(ψi|ψi−1,α)f(d1|ψi(1),α),
Time
Updates
d_3d_1 d_2
![Page 22: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/22.jpg)
EnKS solution
-20
-15
-10
-5
0
5
10
15
20
0 5 10 15 20 25 30 35 40
EnKS: Estimate
0
2
4
6
8
10
0 5 10 15 20 25 30 35 40
EnKS: Errors
![Page 23: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/23.jpg)
EnKS summary
ES and EnKS give identical results for linear models.
EnKS is superior to the ES with nonlinear models.Sequential processing of measurements introduces“Gaussianity”.Ensemble is kept close to the true state.
![Page 24: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/24.jpg)
EnKF: Ensemble Kalman Filter
Special case of EnKS.
Only update state at data times
EnKF forms a prior for the EnKS.
A prediction from EnKF and EnKS will be identical.
Time
Updates
d_3d_1 d_2
![Page 25: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/25.jpg)
EnKF solution
-20
-15
-10
-5
0
5
10
15
20
0 5 10 15 20 25 30 35 40
EnKF: Estimate
0
2
4
6
8
10
0 5 10 15 20 25 30 35 40
EnKF: Errors
![Page 26: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/26.jpg)
Example
Scalar model
∂ψ
∂t= 1 − α+ q,
ψ(t = 0) = 3 + a,
α = 0 + α′,
M(ψ) = d+ ε.
True parameter value is α = 1.
Model operator is linear and independent of ψ.
Solved using EnKF, EnKS and Representer methods.
Exponential time correlation for model errors.
![Page 27: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/27.jpg)
State and parameter estimation
Time
Sol
utio
n
0 10 20 30 40 50
-4
-2
0
2
4
6
8
10
12
14 ObservationsEnKS solRepresenter solEnKS std dev
![Page 28: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/28.jpg)
Estimate of parameter
Time
Par
amet
eres
timat
e
0 10 20 30 40 50
-2
-1
0
1
2
EnKS alphaEnKF alphaEnKF std devEnKS std dev
![Page 29: The combined parameter and state estimation problem · The combined parameter and state estimation problem Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway,](https://reader030.vdocument.in/reader030/viewer/2022040912/5e86a60058f7f502e224eae8/html5/thumbnails/29.jpg)
Summary
www.nersc.no/∼geir/EnKF
Joint state and parameter estimation problem!
Bayesian formulation:Variational methods (multiple minima; MLH).
Hard to compute error statistics!Do not allow for sequential processing ofmeasurements!
Ensemble methods (Gaussianity assumption; mean).Provide estimate with error statistics.Allow for sequential processing of measuremets.· Introduces Gaussianity!· Advantage with nonlinear dynamics.