Simulation of error cycling

Transcription

1 Simulation of error cycling Loïk BERRE, Météo-France/CNRS ISDA, Reading, 21 July 2016 with inputs from R. El Ouaraini, L. Raynaud, G. Desroziers, C. Fischer

2 Motivations and questions EDA and innovations for diagnosing contributions (background errors, observation errors, model errors) in error cycling over one week. Revisit formalism & some previous EDA experiments in the litterature : Respective global amplitudes of these 3 error sources? Evolution of error contributions during the cycling?

3 Contents EDA context at Météo-France Quasi-linear expansion of forecast errors Old background errors / Recent observation errors Observation errors / Model errors

4 Contents EDA context at Météo-France Quasi-linear expansion of forecast errors Old background errors / Recent observation errors Observation errors / Model errors

5 Ensemble of perturbed Data Assimilations (EDA) : simulation of error cycling ε b ε f = Mε a + ε m ε a = (I-KH)ε b + Kε o ε o t i -6h t i t i +6h (e.g. Houtekamer et al 1996, Fisher 2003, Berre et al 2006)

6 Global operational EDA at Météo-France 25 members, T479 (40 km) L105, Arpege 4D-Var (minim T149), 6h cycle. 4D-Var analysis perturbations : observation perturbations (drawn from R, incl. spatial corr. for AMVs), background perturbations (evolved analysis pertbs and model pertbs). Multiplicative inflation of forecast perturbations, using innovation-based diagnostics. Spatially filtered variances for observation QC and minimisation, wavelet-filtered 3D correlations. An EDA is also being developed at mesoscale (AROME, oper 2018).

7 Dynamics of background error variances with EDA Standard deviations of surface pressure (hpa) (2/2/2010) (superimposed with MSLP analysis, in hpa). (e.g. Berre et al 2007, Raynaud et al 2012)

8 Dynamics of horizontal correlations from ensemble and wavelet filtering Length scales (km) for wind near 500 hpa (28/2/2010), superimposed with geopotential (Berre, Varella et Desroziers 2015)

9 Titre de la diapo Dynamics of vertical correlations from ensemble and wavelet filtering Vertical correlations of temperature between 850 & 870 hpa (28/2/2010) (Berre, Varella and Desroziers 2015)

10 Contents EDA context at Météo-France Quasi-linear expansion of forecast errors Old background errors / Recent observation errors Observation errors / Model errors

11 Expansion of forecast error contributions (quasi-linear framework) At a given step t 0 of the cycling : ε a 0 = (I-K 0 H 0 ) ε b 0 + K 0 εo 0 ε f 0 = M 0 εa 0 + εm 0 = M 0 (I-K 0 H 0 ) ε b 0 + M 0 K 0 εo 0 + εm 0 At a later step t i (e.g. one week later) : ε f i = T i+1 ε b 0 + Σ T i-j M j K j εo j + Σ T i-j ε m j j 0 j 0 where T i-j = Π M k (I-K k H k ) = cycling operator k (over k successive analysis/forecast steps from t j to t i ). (El Ouaraini and Berre 2011)

12 Age of error contributions Error contributions over one week, from t 0 to t i : OLD CYCLES t 0 RECENT CYCLES t j t i date of cycle (t) ε b 0 old background errors ε o j εm j recent obs. errors recent model errors ε f i Full expansion of forecast errors at t i : ε f i = T i+1 ε b 0 + Σ T i-j M j K j εo j + Σ T i-j ε m j where T i-j = Π M k (I-K k H k ) = cycling operator k j 0 j 0 (over k successive analysis/forecast steps from t j to t i ). (El Ouaraini and Berre 2011)

13 Several processes in error cycling Expansion of forecast errors : ε f i = T i+1 ε b 0 + Σ T i-j M j K j εo j + Σ T i-j ε m j where T i-j = Π M k (I-K k H k ) = cycling operator. k j 0 j 0 Old background errors ε b 0 : repeted analysis damping & model propagation. Recent observation errors ε o j : filtering (K) & propagation (M) ; damping & propagation (T) ; accumulation (Σ). Recent model errors ε m j : damping & propagation ; accumulation.

14 Links with EDA and innovations Same equation for errors and perturbations : ε f i = T i+1 ε b 0 + Σ T i-j M j K j εo j + Σ T i-j ε m j j 0 j 0 Simulation of error cycling : cycle observation and model perturbations added to deterministic system (e.g. with 4D-Var and non linear forecasts included). Observation and model perturbations are often (+/- indirectly) derived from innovation-based estimates of covariances. Amplitudes of some estimated (accumulated) error contributions can be diagnosed and compared using EDA and innovations.

15 Experimental framework Revisit some sensitivity experiments with a previous EDA configuration : no model perturbations (offline tuning of horizontally averaged variances) ; same K in all compared configurations (static spectral correlations, flow-dependent variances). Comparison «cold start» / «warm start» ensembles : diagnosis / evolution of old background perturbations. EDA configurations without (recent) model perturbations : to diagnose the 2 other contributions (old background errors, recent observation errors). to estimate model error variance by comparison with innovation-based estimates of forecast errors. Offline diagnosis of global amplitudes (variances) of perturbations.

16 Contents EDA context at Météo-France Quasi-linear expansion of forecast errors Old background errors / Recent observation errors Observation errors / Model errors

17 Contribution of recent observation errors ε f i = T i+1 ε b 0 + Σ T i-j M j K j εo j + Σ T i-j ε m j j 0 j 0 Contribution of recent observation errors to forecast errors : ε fo i = Σ T i-j M j K j ε o j j 0 Simulation and evolution using EDA : * use ε b 0 =0 (cold start = start from unperturbed background), then cycle with ε o j ( j 0 ) (and with εm j = 0). * compute evolution of EDA spread : σ( ε fo t = Σ T t-j M t K t ε o t ).

18 Accumulation of recent observation errors Evolution of σ( ε fo t = Σ T t-j M j K j ε o j ) j 0 at successive analysis steps (t 0, t 1, t i ) σ(ε fo t ) (vorticity (10-5 s -1 ), 500 hpa) (El Ouaraini and Berre 2011, see also Fisher et al 2005) Observation error contribution increases (accumulation), then it stabilizes after 3-5 days of cycling, due to analysis damping : T t-j = Π M k (I-K k H k ). k

19 Accumulation of recent observation errors Evolution of ε fo t = Σ T t-j M j K j ε o j j 0 at successive analysis steps (t 0, t 1, t 2, t i ) : ε fo 0 = M 0 K 0 εo 0 ε fo 1 = M 1 K 1 εo 1 + T 1 εfo 0 ε fo 2 = M 2 K 2 εo 2 + T 1 εfo 1 + T 2 εfo 0. ε fo t = M t K t ε o t + T 1 ε fo t-1 + T 2 εfo t T 12 εfo t-12 + Σ T k εfo t-k k>12 T is mainly damping, so old obs. error contrib. (with time-lag k>12) become negligible, and σ(ε fo t ) stabilizes ~ at step t = 12 (3 days).

20 Contribution of old background errors ε f i = T i+1 ε b 0 + Σ T i-j M j K j εo j + Σ T i-j ε m j Contribution of «old» background errors : ε fb i = T i+1 εb 0 with T i+1 = Π M k (I-K k H k ) Simulation and evolution using EDA : use warm start = ensemble started 6 days before t 0 : * use ε b 0 =Σ T 0-j M j K j εo j (= contribution of old observation errors), j <0 j 0 i k=0 j 0 then cycle with ε o j ( j 0 ) (and with εm j = 0). * compute evolution of sqrt of σ²(ε fb t ) = σ²(εfb t + εfo t ) σ²(εfo t ) (~sqrt of EDA variance difference between warm start and cold start)

21 Evolution of old background error contribution σ(ε fb t = T t+1 εb 0 ) t with T t+1 = Π M k (I-K k H k ) k=0 σ(ε fo t ) (vorticity (10-5 s -1 ), 500 hpa) ε b 0 = Σ T 0-j M j K j εo j (contribution of old observation errors) j <0 σ(ε fb 0 ) > σ(εfo 0 ) (for vorticity). σ(ε fb i = T i+1 εb 0 ) decreases ~ linearly / analysis damping.

22 Contribution of old+recent observation errors σ(ε fb t + εfo t ) σ(ε fb t = T t+1 εb 0 ) σ(ε fo t ) (vorticity (10-5 s -1 ), 500 hpa) Total contribution of (old+recent) obs errors is stable : damping of old observation errors is compensated by accumulation of recent obs errors.

23 Dependence on amplitude of old background errors σ(ε fb t = T t+1 εb 0 ) (vorticity (10-5 s -1 ), 500 hpa) ε b 0 = Σ T 0-j M j K j εo j (old observation errors) j <0 ε b 0 = B1/2 η (old model errors + old observation errors) σ(ε fb t = T t+1 εb 0 ) decreases ~ linearly / analysis damping.

24 First conclusions / formalism + experiments Quasi-linear expansion of forecast errors / contributions of old background errors, recent observation errors, recent model errors. Compare cold/warm start EDA (without recent model perturbations) to diagnose error contrib. of old background and recent observations. Old background errors vanish (~linearly) after 3-5 days of cycling. Compensated by accumulation of recent observation errors. This may help for interpretation/diagnosis of total forecast errors (~ recent observation & model errors).

25 Contents EDA context at Météo-France Quasi-linear expansion of forecast errors Old background errors / Recent observation errors Observation errors / Model errors

26 Innovation covariances background error covariance E[ (y Hx b )(y Hx b ) T ] = R + HBH T separation

27 Covariances of analysis residuals y δy = y H(x b ) (innovation) δy H δx = H(x b ) H(x a ) (increment) ε o E[H δx δy T ] = HB*H T δx x a ε a E[(y H(x a )) δy T ] = R* x b ε b x* (true state) (Desroziers et al 2005)

28 Estimation of model error contributions (to forecast errors) t 0 t j t i date of cycle ε b 0 ε o j εm j ε f i Error ε f i of forecast issued from cycle t i : ε f i = T i+1 ε b 0 + Σ T i-j M j K j εo j + Σ T i-j ε m j [ T i-j = Π M k (I-K k H k ) ] j 0 j 0 k obs. errors model errors accumulated from t 0 to t i Compare σ δy ( εf i = ε fb0 i + Σ T i-j M j K j εo j + Σ T i-j ε m j ) (innovation-based) with σ EDA ( ε fb0 i + Σ T i-j M j K j εo j ) (using unperturbed model from t 0 to t i ).

29 Total forecast error versus observation error contribution σ δy (εf i = ε fb0 i +Σ T i-j M j K j εo j + Σ T i-j εm j ) σ EDA ( ε fb0 i +Σ T i-j M j K j εo j ) Pressure (hpa) Standard deviation of forecast errors (aircraft observations of temperature (K)) (Raynaud et al 2012)

30 Total forecast error versus observation error contribution σ 2 δy ( εfb0 i +Σ T i-j M j K j εo j + Σ T i-j εm j ) ~ σ2 ( Σ T i-j M j K j ε o j ) + σ2 ( Σ T i-j ε m j ) If cov(ε fb0 i, εfm i ) ~ 0, i.e. if the one-week period is large enough to neglect time correlations between old background errors & recently accumulated model errors, in addition to : σ 2 EDA ( εfb0 i +Σ T i-j M j K j εo j ) = σ2 EDA ( Σ T i-j M j K j εo j ) if σ2 (ε fb0 i ) ~ 0

31 Quantification of model error accumulated during cycling σ δy/eda ( εfm i = Σ T i-j ε m j ) σ EDA ( ε fo i = Σ T i-j M j K j ε o j ) Pressure (hpa) Standard deviation of error contributions aircraft observations of temperature (K))

32 Possible guidance for model error representation? Variance of model error accumulated over 3-5 days can be diagnosed : model error contrib. ~ one half of 6h forecast error variance ; obs. error contrib. ~ other half. Variance of model error accumulated over 6h: similar formalism, but assumptions/diagnostics on temporal correlations required. Importance of innovation-based estimates (R and B). Possible model error representations (Q 1/2 η, SPPT, SKEB, etc) may be compared/adjusted with such diagnostics.

33 Conclusions EDA for estimation of flow-dependent covariances, combined with spatial filtering methods. EDA (and innovations) for diagnosing contributions to error cycling : obs. errors contribute ~ within the last 3-5 days of DA cycling. old background errors vanish ~ after 3-5 days of DA cycling. model error contrib. are similar in amplitude to obs. error contrib. LAM : contribution of LBC errors over 1/3 of ALADIN-France domain due to advection in the cycling (El Ouaraini et al 2015). Extension to diagnostics of spatial structures, pursue comparison EDA vs innovations, etc.

34 Thank you for your attention