A Local Ensemble Transform Kalman Particle Filter
Sylvain Robert and Hans R. KünschETH Zürich, Seminar for Statistics
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 1
Outline
• Introduction
• Algorithm
• Numerical experiments
• Conclusions
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 2
Motivation
• Target application: Convective Scale Data Assimilation
• Challenges:
– high resolution (≈1km)
– non-linear forecasting step xat → xb
t+1.
– non-Gaussian background distribution πbt (x).
– non-Gaussian analysis distribution πat (x) (even if likelihood is
linear and Gaussian).
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 3
Introduction
• We assume that y |x ∼ N (Hx , R) and skip time index t.
• xb and xa: background and analysis ensembles.
• πb(x) and πa(x): background and analysis distributions.
Analysis:
• xb "+" y → xa
• Bayes’ formula: πa(x) ∝ πb(x) · `(y |x)
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 4
Introduction
different assumptions = different solutions
• πb(x) Gaussian + unlimited computation→ Kalman Filter (KF)
• πb(x) non-Gaussian + unlimited computation→ Particle Filter (PF)
• πb(x) Gaussian + limited computation→ EnKF
• πb(x) non-Gaussian + limited computation→ ???
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 5
The EnKPF in a nutshell
Frei and Künsch (Biometrika, 2013).
• πa(x) ∝ πb(x) · `(y |x) = πb(x) · `(y |x)γ︸ ︷︷ ︸∝πγ (x)
·`(y |x)1−γ
• Two steps:πb(x) EnKF−−−−−→
γπγ(x) PF−−−−−→
1−γπa(x)
• γ = 1→ EnKF
• γ = 0→ PF
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 6
The EnKPF in a nutshell
●
●
●
●
●
Analysis distribution πaEnKPF (x):
xa,j ∼k∑
i=1
αγ,i N (µγ,i , Pa,γ)
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 7
Outline
• Introduction
• Algorithm
– Ensemble space
– Localization
• Numerical experiments
• Conclusions
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 8
Ensemble space
• Split ensembles into mean and deviations:
xb = x̄b1′ + X b and xa = x̄a1′ + X a
• Use the empirical covariance Pb = 1k−1X b(X b)′
• Analysis mean :
x̄a = x̄b + X bm, m = k × 1
• Analysis deviations:
X a = X bW , W = k × k
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 9
EnKPF in ensemble space
m = mµ + 1k W µW α1
W = W µW α + W ε − 1k W µW α11′
• mµ and W µ: mixture components µγ,i
• W α: particle resampling
• 1k W µW α1: correction of the mean due to resampling.
• W ε: individual perturbations to ensure correct covariance:
– stochastic
– deterministic: transform filter
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 10
Step 1: analysis mean
●
●
First we move the ensemble mean
towards the observation:
x̄a = x̄b + X bmµ
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 11
Step 2: mixture components
(x̄b + X bmµ)1′ + X bW µ
●
●
0.0
0.2
0.4
0.6
0.8
1.0
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 12
Step 3: weights and resampling
(x̄b + X bmµ)1′ + X bW µWα
●
1
1
3
0.0
0.2
0.4
0.6
0.8
1.0
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 13
Step 4: individual perturbations
X bW ε ∼ N (0, Pa,γ), Transform version
0.0
0.2
0.4
0.6
0.8
1.0
Remark: W ε is not just a square-root of Pa,γ → solve for Cov (xa).
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 14
All together
(x̄b + X bmµ)1′ + X bW µWα + W ε
●
●
●
●
●
0.0
0.2
0.4
0.6
0.8
1.0
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 15
LocalizationThe curse of dimensionality:
• PF: necessary number of particles increases exponentially.
• EnKPF: a bit better but not immune to the problem.
Possible remedies:
• Carefully chosen proposal distribution.
• Localization.
Problem:
• Not easy to apply to PF methods.
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 16
Localization: step 1
Local weights, resample locally and glue together→ discontinuities
longitude
field
val
ue
Each line is an analysis particle with three cases highlighted in color.
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 17
Localization: step 2
Permute indices locally→ remove some discontinuities.
longitude
field
val
ue
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 18
Localization: step 3
Interpolate W on finer grid→ smooth out remaining discontinuities.
longitude
field
val
ue
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 19
Summary of new algorithm
Local Ensemble Transform Kalman Particle Filter: LETKPF
• Hybrid between EnKF and PF: handles non-Gaussian distributionswhile maintaining sample diversity.
• Transform: guarantees exact covariance with deterministic scheme.
• Local: uses local weights and local resampling while avoiding prob-lems with discontinuities.
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 20
Outline
• Introduction
• Algorithm
• Numerical experiments: COSMO-KENDA
– Case study (7th of June 2015 12 UTC)
– Cycled experiment (June 04-16)
– Forecast experiment (12 hours)
• Conclusions
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 21
COSMO-KENDA experiment
In collaboration with Daniel Leuenberger from Meteoswiss and with thehelpful support of DWD.
• Area surrounding Switzerland, high resolution (≈ 2.2km).
• 04-16th of June 2015 (period of intense convective activity).
• 40 ensemble members.
• Assimilation of conventional observations (no radar).
• Algorithms: LETKPF and LETKF.
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 22
7th of June 12 UTCZonal and meridional wind components: ensemble mean and std.
U
V
−15
−10
−5
0
5
10
15
MeanU
V
1
2
3
4
Std
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 23
Local particle weights: αγ,i
• Different particles fit the observations better in different places.
• 1/40 < α < 2/40→ particle resampled once or twice.
• 2/40 < α < 3/40→ particle resampled twice or thrice.
• . . .
1 2 3
12345678
1/40
α
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 24
Combination of particles
(xa,1 − x̄a) = (xb,1 − x̄b)W11 + (xb,2 − x̄b)W21 + (xb,3 − x̄b)W31 + ...
1 2 3
12345678
1/40
α
1 2 3
0
1
Wi1
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 25
Discontinuities
• Contribution of particle 1 to its own analysis: W11.
• Large continuous patches: good.
• Some discontinuous patterns could be fixed, but it would requiresome global communication (work in progress).
12345678
1/40
0
1
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 26
Adaptive choice of γ
• Chosen locally such that ESS=50% (≈ half of the mixture compo-nents µγ,i are used).
• Small γ means more PF, big γ more EnKF.
• Joint property of the background distribution and the observations.
0
0.5
1value
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 27
Cycled experiment
• Hourly assimilation of conventional observations.
• LETKPF vs LETKF.
• Vertical profiles of RMSE and spread for T, RH and WIND.
• Averaged over whole period 04-16.06.2015.
Parameters:
• Parameter γ chosen adaptively such that ESS=50%.
• Localization radius: chosen adaptively.
• Multiplicative covariance inflation.
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 28
Wind
300
400
500
600
700
850
925
1000
0.0 0.5 1.0 1.5 2.0 2.5error
hPa
RMSE
300
400
500
600
700
850
925
1000
0.0 0.5 1.0 1.5value
Spread
300
400
500
600
700
850
925
1000
−0.1 0.0 0.1 0.2 0.3 0.4 0.5error
typeanalysis
forecast
modelLEnKPF
LETKF
LETKPF
Bias
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 29
Temperature
300
400
500
600
700
850
925
1000
0.0 0.5 1.0 1.5error
hPa
RMSE
300
400
500
600
700
850
925
1000
0.0 0.1 0.2 0.3 0.4 0.5value
Spread
300
400
500
600
700
850
925
1000
−0.1 0.0 0.1 0.2 0.3error
typeanalysis
forecast
modelLEnKPF
LETKF
LETKPF
Bias
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 30
Relative humidity
300
400
500
600
700
850
925
1000
0.00 0.05 0.10 0.15 0.20error
hPa
RMSE
300
400
500
600
700
850
925
1000
0.00 0.02 0.04 0.06value
Spread
300
400
500
600
700
850
925
1000
−0.075−0.050−0.025 0.000 0.025 0.050error
typeanalysis
forecast
modelLEnKPF
LETKF
LETKPF
Bias
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 31
Outline• Introduction
• Algorithm
• Numerical experiments
– Case study (7th of June 2015 12 UTC)
– Cycled experiment (June 04-16)
– Forecast experiment (12 hours)• Conclusions
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 32
Forecast: RMSE and bias
bias
PS
bias
WIND
rmse
PS
rmse
WIND
0
20
40
60
0.00
0.05
0.10
0.15
60
80
100
1.7
1.8
1.9
2.0
0 200 400 600 200 400 600
0 200 400 600 200 400 600lead time (min)
erro
r methodLETKF
LETKPF
SYNOP
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 33
Forecast: ETS of TOT_PREC
How well did the forecast predict rain > 0.1 mm/h (accounting for chance)?
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 34
Forecast: FBIFrequency Bias Index: 1: perfect, >1: over-, <1: under-forecasting.
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 35
Conclusions• New algorithm: LETKPF
– Hybrid EnKF and PF in ensemble space.
– transform and local.
• Positive results with COSMO-KENDA:
– Case study of 07.06.2015: reasonable behavior.
– Transform filter better than stochastic version.
– Results equivalent to LETKF on cycled experiment.
– Successful 12 hour forecasts.
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 36
Thank you!
Questions?
Seminar for StatisticsSfS Sylvain Robert and Hans R. Künsch 18. July 2016 37