Practical Aspects of Ensemble-based Kalman Filters

Transcription

1 Practical Aspects of Ensemble-based Kalman Filters Lars Nerger Alfred Wegener Institute for Polar and Marine Research Bremerhaven, Germany and Bremen Supercomputing Competence Center BremHLR Bremen, Germany

2 Outline Aspects Computing Analysis formulation Localization Collaborations: AWI: W. Hiller, J. Schröter, S. Loza, P. Kirchgessner, T. Janjic (now DWD) BSH: F. Janssen, S. Massmann Bremen University: A. Bunse-Gerstner

3 The Problem

4 Application Example Model surface temperature Satellite surface temperature Norway Sweden Finland UK Oberwolfach Information: Model Information: Observation Forecasting in North & Baltic Seas Combine model and observations for optimal initial condition State vector size: (4 fields 3D, field 2D) Obervations: (Surface temperature only) Ensemble size 8 S. Loza et al., Journal of Marine Systems 05 (202) 52-62

5 Forecast deviation from satellite data No assimilation Assimilation RMS bias Improvements also sub-surface and in other fields

6 Data Assimilation Problem: Estimate model state (trajectory) from guess at initial time model dynamics observational data Characteristics of system: approximated by discretized differential equations high-dimension - O( ) sparse observations non-linear Current standard methods: Optimization algorithms ( 4DVar ) This talk! Ensemble-based estimation algorithms

7 Ensemble-based Kalman Filter First formulated by G. Evensen (EnKF, 994) Kalman filter: express probability distributions by mean and covariance matrix EnKF: Use ensembles to represent probability distributions state estimate ensemble forecast initial sampling forecast analysis ensemble transformation observation Looks trivial! BUT: There are many possible choices! time 0 time time 2

8 Computational and Practical Issues Data assimilation with ensemble-based Kalman filters is costly! Memory: Huge amount of memory required (model fields and ensemble matrix) Computing: Huge requirement of computing time (ensemble integrations) Parallelism: Natural parallelism of ensemble integration exists (needs to be implemented) Fixes : Filter algorithms do not work in their pure form ( fixes and tuning are needed) because Kalman filter optimal only in linear case Lars Nerger Ensemble square-root KFs

9 What we are looking for Goal: Find the assimilation method with smallest estimation error most accurate error estimate least computational cost least tuning Want to understand and improve performance Difficulty: Optimality of Kalman filter well known for linear systems No optimality for non-linear systems limited analytical possibilities apply methods to test problems Lars Nerger Ensemble square-root KFs

10 Computing

11 Logical separation of assimilation system single program time state Filter Initialization analysis re-initialization Core of PDAF state observations Model initialization time integration post processing mesh data Observations obs. vector obs. operator obs. error Explicit interface Indirect exchange (module/common) Nerger, L., Hiller, W. (202). Software for Ensemble-based DA Systems Implementation and Scalability. Computers and Geosciences. In press. doi:0.06/j.cageo

12 PDAF: A tool for data assimilation PDAF - Parallel Data Assimilation Framework a software to provide assimilation methods an environment for ensemble assimilation for testing algorithms and real applications useable with virtually any numerical model also: apply identical methods to different models test influence of different observations makes good use of supercomputers (Fortran and MPI; tested on up to 4800 processors) More information and source code available at

13 Analysis Formulations

14 Ensemble-based/error-subspace Kalman filters A little zoo (not complete): New study ESTKF _ Properties and (Nerger differences et al. 202) are hardly understood _ EnKF(94/98) Studied in Nerger et al. (2005) RRSQRT ROEK SEEK SEIK EnKF(2003) EnKF(2004) EAKF EnSRF ETKF Learn from studying relations and differences MLEF SPKF ESSE RHF anamorphosis New filter formulation ESTKF: L. Nerger et al., Monthly Weather Review 40 (202)

15 Model Equations Stochastic dynamic model: x t i = M i,i [x t i ]+ i, x t i, i R n Stochastic observation model: y k = H k [x t k]+ k, y k, k R m i N(0, Q i ); i T j = Q i ij k N (0, R k ); k T l = R k kl Assumptions: x t N i N ( x t i, P i ) k T k =0; i (x ti) T =0; k (x tk) T =0 Model error Observation error

16 The Ensemble Kalman Filter (EnKF, Evensen 94) Initialization: Generate random ensemble Ensemble statistics approximate and covariance Forecast: x a(l) i = M i,i [x a(l) (l) i ]+ i Analysis: x a(l) k = x f(l) k + K k y (l) k H k x f(l) k Kalman filter K k = P f k HT k H k P fk HTk + R k P f k := N T x f(l) k x f k x f(l) k x f k N l= x a k := N x a(l) k N l=

17 Issues of the EnKF94 Monte Carlo Method ensemble of observations required (samples matrix R; introduces sampling error) Inversion of large matrix H k P f k HT k + R k R m m (can be singular, possibly large differences in eigenvalues >0) Alternative: Compute analysis in space spanned by ensemble Methods: Ensemble Square-Root Kalman Filters, e.g. SEIK (Pham et al., 998) ETKF (Bishop et al., 200)

18 Ensemble Transform Kalman Filter - ETKF Ensemble perturbation matrix X k := X k X k size (n x N) Analysis covariance matrix P a = X f A(X f ) T. Transform matrix (in ensemble space) A := (N )I +(HX f ) T R HX f. Ensemble transformation X a = X f W ETKF. Ensemble weight matrix W ETKF := N CΛ. (n x n) (N x N) (n x N) (N x N) CC T = A (symmetric square root) Λ is identity or random orthogonal matrix with EV (,...,) T )

19 SEIK Filter Error-subspace basis matrix L := X f T, size (n x N-) (T subtracts ensemble mean and removes last column) Analysis covariance matrix P a = LÃLT (n x n) Transform matrix (in error subspace) à := (N )T T T +(HL) T R HL. (N- x N-) Ensemble transformation X a = LW SEIK. (n x N) Ensemble weight matrix W SEIK := N CΩ T (N- x N) C. is square root of à (originally Cholesky decomposition) CΩ T is transformation from N- to N (random or deterministic)

20 Weight Matrices (W in X a = X f W ) ETKF ETKF: Transformation matrix 0. SEIK-Cholesky sqrt SEIK chol: Transformation matrix ETKF main contribution from diagonal (minimum transformation) Off-diagonals of similar weight Minimum change in distribution of ensemble variance SEIK with Cholesky sqrt main contribution from diagonal Off-diagonals with strongly varying weights Changes distribution of variance in ensemble

21 Transformation Matrix of SEIK/symmetric sqrt SEIK symmetric sqrt SEIK sym: Transformation matrix Difference SEIK-ETKF transformation matrices difference: SEIK ETKF x Transformation matrices of ETKF and SEIK-sym very similar Largest difference for last ensemble member (Experiments with Lorenz96 model: This can lead to smaller ensemble variance of this member)

22 SEIK depends on ensemble order Switch last two ensemble members SEIK sym: Difference of transformation matrices x (Switched back last two columns & rows for comparison) Ensemble transformation depends on order of ensemble members (For ETKF the difference is 0-5 ) Statistically fine, but not desirable!

23 Revised T matrix Identical transformations require different projection matrix for SEIK: For SEIK: T subtracts ensemble mean and drops last column Dependence on order of ensemble members! Solution: Redefine T: Distribute last member over first N- columns Also replace Ω L := X f T, by new ˆT New filter formulation: Error Subspace Transform Kalman Filter (ESTKF)

24 T-matrix in SEIK and ESTKF SEIK: ESTKF: T i,j = ˆT i,j = N N N N N p + N p + N p N for i = j, i < N for i = j, i < N for i = N for i = j, i < N for i = j, i < N for i = N Efficient implementation as subtraction of means & last column ETKF: improve compute performance using a matrix T

25 ESTKF: New filter with identical transformation as ETKF New filter ESTKF properties like ETKF: Minimum transformation Transformation independent of ensemble order But: analysis computed in dimension N- direct access to error subspace smaller condition number of A L. Nerger et al., Monthly Weather Review 40 (202)

26 Localization

27 Localization: Why and how? Combination of observations and model state based on estimated error covariance matrices Example: Sampling error and localization true sampled localized Finite ensemble size leads to significant sampling errors particularly for small covariances! distance Remove estimated long-range correlations Increases degrees of freedom for analysis (globally not locally!) Increases size of analysis correction

28 Global vs. local SEIK, N=32 (March 993) Error reduced to 83.6% Error reduced to 3.7% Improvement is error reduction by assimilation Localization extents improvements into regions not improved by global SEIK Regions with error increase diminished for local SEIK Underestimation of errors reduced by localization L. Nerger et al. Ocean Dynamics 56 (2006) 634

29 Localization Types Simplified analysis equation: x a = x f + Pf P f + R y xf Covariance localization Modify covariances in forecast covariance matrix P f Element-wise product with correlation matrix of compact support Requires that P f is computed (not in ETKF or SEIK) E.g.: Houtekamer/Mitchell (998, 200), Whitaker/Hamill (2002), Keppenne/ Rienecker (2002) Observation localization Modify observation error covariance matrix R Needs distance of observation (achieved by local analysis or domain localization) Possible in all filter formulations E.g.: Evensen (2003), Ott et al. (2004), Nerger/Gregg (2007), Hunt et al. (2007)

30 Local SEIK filter domain & observation localization Local Analysis: Update small regions (like single vertical columns) Observation localizations: Observations weighted according to distance Consider only observations with weight >0 State update and ensemble transformation fully local S: Analysis region D: Corresponding data region Similar to localization in LETKF (e.g. Hunt et al, 2007) L. Nerger et al., Ocean Dynamics 56 (2006) 634 L. Nerger & W.W. Gregg, J. Mar. Syst. 68 (2007) 237

31 Different effect of localization methods Experimental result: Twin experiment with simple Lorenz96 model Covariance localization better than observation localization (Also reported by Greybush et al. (20) with other model) Time-mean RMS errors / Inflation forgetting factor Covariance EnKF sqrt, obs. localization error= support radius Localization radius / inflation T. Janjic et al., Mon. Wea. Rev. 39 (20) forgetting factor Observation LSEIK fix, obs. localization error= support radius Localization radius

32 Different effect of localization methods (cont.) Larger differences for smaller observation errors / Inflation forgetting factor Covariance localization EnKF sqrt, obs. error= support radius Localization radius forgetting factor Observation LSEIK fix, obs. localization error= support radius Localization radius

33 Covariance vs. Observation Localization Some published findings: Both methods are similar Slightly smaller width required for observation localization But note for observation localization: Effective localization length depends on errors of state and observations Small observation error wide localization Possibly problematic: in initial transient phase of assimilation if large state errors are estimated locally effective weight effective weight distance prescribed weight effective weight P= R= P= R= distance P: state error variance R: observation error variance

34 Regulated Localization New localization function for observation localization formulated to keep effective length constant (exact for single w OLR = wcl σr 2 observation) ( HPH T + σr 2 wcl HPH T ) depends on state and observation HPH T + errors σr 2 depends on fixed localization function cheap to compute for each observation Only exact for single observation works for multiple weight w fixed w reg, R=0 w reg, R=.0 w reg, R=0. P= distance L. Nerger et al. QJ Royal. Meterol. Soc. 38 (202)

35 Lorenz96 Experiment: Regulated Localization forgetting factor Fixed Covariance localization, loc., N=0, R= EnKF sqrt, LSEIK fix, obs. error= support radius Reduced minimum rms errors Increased stability region forgetting factor Still need to test in real application Regulated localization, N=0, R= LSEIK reg, obs. error= support radius Description of effective localization length explains the findings of other studies!

36 Summary Ensemble-based KFs not exact But they work! Improve methods Least cost; least tuning; best state and error estimates Study relations for improvements Efficient analysis formulations Efficient localization Thank you!