1. ISDA 2014, Feb 24 – 28, Munich 1 Impact of ensemble perturbations provided by convective-scale ensemble data assimilation in the COSMO-DE model Florian Harnisch 1,Christian Keil 2 1 Hans-Ertel-Centre for Weather Research, Data Assimilation, LMU München, Germany 2 Meteorologisches Institut, LMU München, Germany Special thanks to Hendrik Reich & Andreas Rhodin, DWD

2. ISDA 2014, Feb 24 – 28, Munich 2 KENDA-COSMO ensemble of COSMO-DE first-guess forecasts + set of observations → ensemble of analyses → ensemble of high-resolution initial conditions to directly(?) initialise ensemble forecasts Kilometer-Scale Ensemble Data Assimilation (KENDA) → Lokal Ensemble Transform Kalman Filter (LETKF) (Hunt el al. 2007) applied for the COSMO-DE model

3. ISDA 2014, Feb 24 – 28, Munich 3 KENDA-COSMO: Inflation  LETKF: background error covariance matrix P b is estimated from ensemble forecasts x b Problem: not all sources of forecast error are sampled in P b → sampling errors due to limited ensemble size & model error → estimate of P b will systematically underestimate variances Solution: Inflation of estimate of P b to enhance the variance (1) multiplicative covariance inflation (adaptive / fixed) (2) relaxation-to-prior-perturbations / relaxation-to-prior-spread (Zhang et al. 2004)(Whitaker and Hamill, 2012) =

4. ISDA 2014, Feb 24 – 28, Munich 4 Setup of experiments KENDA: - 3-hourly LETKF data assimilation of conventional data - 3-hourly analysis ensemble with 20 ensemble members - 20 member ECMWF EPS lateral boundary conditions (16 km) - No physics parametrization perturbations (PPP) - Multiplicative adaptive covariance inflation KENDAppp: including 10 physics parametrization perturbations (PPP) KENDArtpp: relaxation-to-prior-perturbation inflation (α = 0.75 ) KENDArtps: relaxation-to-prior-spread inflation (α= 0.95 ) KENDArtps40: 40 ensemble members / relaxation-to-prior-spread (1) 15 UTC 10 June - 00 UTC 12 June 2012: → 21-h fc at 00 UTC 11 / 12 June (2) 06 UTC 18 June – 12 UTC 19 June 2012: → 21-h fc at 12 UTC 18 June

5. ISDA 2014, Feb 24 – 28, Munich 5 KENDA covariance inflation, 12 UTC 11 June 2012 Analysis ensemble spread U-Wind (m s -1 ) Radar derived precipitation (mm/h) First-guess ensemble spread U-Wind (m s -1 ) Observation used in the LETKF data assimilation

6. ISDA 2014, Feb 24 – 28, Munich 6 KENDA relaxation-to-prior-pert, 12 UTC 11 June 2012 Analysis ensemble spread U-Wind (m s -1 ) Radar derived precipitation (mm/h) First-guess ensemble spread U-Wind (m s -1 ) Observation used in the LETKF data assimilation

7. ISDA 2014, Feb 24 – 28, Munich 7 Departure statistics for KENDA experiment  Accuracy of the analysis ensemble mean (solid) compared to the first-guess (+3 h) ensemble mean (dashed) → relaxation method inflation ensemble = better accuracy Radiosonde temperature KENDArtps KENDA N Obs

8. ISDA 2014, Feb 24 – 28, Munich 8 Departure statistics for KENDA experiment  Accuracy of the analysis ensemble mean (solid) compared to the first-guess (+3 h) ensemble mean (dashed) → larger ensemble = better accuracy N Obs Radiosonde temperature KENDArtps KENDArtps40

9. ISDA 2014, Feb 24 – 28, Munich 9 Ensemble mean error and ensemble spread  Average over 11 cycles  Verification against COSMO-DE analysis  PPP increase the spread  Relaxation methods lead to the largest spread (RMSE~SPREAD) +3 h forecast of U-Wind: KENDA KENDAppp KENDArtpp errorspread

10. ISDA 2014, Feb 24 – 28, Munich 10 Ensemble rank histogram Verified against COSMO-DE analysis (similar results against observations ) OPER KENDA +3 h forecasts of 10 m wind speed rank frequency KENDAppp KENDArtps

11. ISDA 2014, Feb 24 – 28, Munich 11 Ensemble dispersion  Normalized variance difference (NVD): var(eps 1) - var(eps 2) var(eps1) + var(eps 2) average all cycles NVD KENDA / OPER KENDA / KENDAppp KENDA / KENDArtps forecast steps (h) 1-h prec KENDA / KENDArtps40

12. ISDA 2014, Feb 24 – 28, Munich 12 BSS: 3-h ensemble forecasts of precipitation  Brier Skill Score = [resolution – reliability] / uncertainty  Hard to beat COSMO-DE-EPS on up to 3-h hours: LHN in analysis  Impact of model physics perturbations, inflation method and ensemble size 15 UTC 10 June – 00 UTC 12 June 2012 06 UTC 18 June – 12 UTC 19 June 2012 BSS thresholds (mm / 3h) KENDA KENDAppp KENDArtps KENDArtps40 OPER KENDA KENDArtps OPER

13. ISDA 2014, Feb 24 – 28, Munich 13 BSS: 21-h ensemble forecasts of precipitation  Brier Skill Score = [resolution – reliability] / uncertainty  Accounting for model errors with PPP shows positive impact  Large impact of inflation procedure 3-21 h forecasts averaged over Germany thresholds (mm / 3h) 00 UTC 11 June 201200 UTC 12 June 2012 KENDA KENDAppp KENDArtps OPER KENDA KENDAppp KENDArtps OPER BSS

14. ISDA 2014, Feb 24 – 28, Munich 14 Summary  KENDA-COSMO ensemble of analyses → Consistent ICs for ensemble forecasts → ICPs are present at all scales / all levels from the beginning  Necessary to use inflation methods to account for unrepresented error sources: large impact of different methods  Physic parameter perturbations can only partially account for model error  Ensemble size matters (initialize 20 member FC from 40 member AN?)

15. ISDA 2014, Feb 24 – 28, Munich 15 Operational COSMO-DE-EPS  Downscaling of 4 global models via intermediate COSMO run (7 km) → perturbations are added to deterministic high-res COSMO DE analysis  5 variables of the model physics parametrizations are perturbed in the FC → 4 * 5 = 20 member ensemble  Initial ensemble perturbations have a clear positive impact in the first 9-12 hours of the forecast Evaluation of summer 2011 from Kühnlein et al. 2014 BSS with / without initial perturbations

16. ISDA 2014, Feb 24 – 28, Munich 16 Ensemble mean error and ensemble spread +3 h forecast of U-Wind at 3.1 km: KENDA OPER KENDArtpp error spread  Verification against COSMO-DE analysis  OPER has smaller error: constructed around COSMO-DE analysis

17. ISDA 2014, Feb 24 – 28, Munich 17 Ensemble mean error and ensemble spread  Average over 11 cycles  Verification against COSMO-DE analysis  OPER has smaller error: → choice of verification → initial ensemble constructed around COSMO-DE analysis +3 h forecast of U-Wind: KENDA OPER KENDArtpp errorspread

18. ISDA 2014, Feb 24 – 28, Munich 18 Ensemble dispersion  Normalized variance difference (NVD): var(eps 1) - var(eps 2) var(eps1) + var(eps 2) average all cycles NVD U (~3.1 km) KENDA / OPER KENDA / KENDAppp KENDA / KENDArtps forecast steps KENDA / KENDArtps40

19. ISDA 2014, Feb 24 – 28, Munich 19 Physics parameter perturbations PPP  Setup follows Gebhardt et al. (2011), Peralta et al. (2012) namelistdefaultperturbed member entr_sc0.00030.0021; 11 q_crit4.01.62; 12 rlam_heat1.00.13; 13 rlam_heat1.0104; 14 tur_len5001505; 15 clc_diag0.50.756; 16 mu_rain0.507; 17 mu_rain0.51.58; 18 cloud_num5.0E+085.0E+079; 19 cloud_num5.0E+085.0E+0910; 20 used in operational setup of COSMO-DE-EPS

20. ISDA 2014, Feb 24 – 28, Munich 20 21-h ensemble forecasts of precipitation fc steps (h) KENDA KENDAppp OPER Ensemble Mean Radar Forecast of 1-h precipitation averaged over Germany, 00 UTC 11 June 2012 KENDArtps fc steps (h)

21. ISDA 2014, Feb 24 – 28, Munich 21 21-h ensemble forecasts of precipitation fc steps (h) Forecast of 1-h precipitation averaged over Germany, 00 UTC 12 June 2012 fc steps (h) KENDA KENDAppp OPER Ensemble Mean Radar KENDArtps

22. ISDA 2014, Feb 24 – 28, Munich 22  improved discrimination KENDA → KENDArtps observed relative freq KENDA / OPER KENDArtps / OPER forecast probability false alarm rate hit rate ~0.86 ~0.85 ~0.86 ~0.81  1 mm / 3h precipitation from 15 UTC 10 June – 00 UTC 12 June 2012  Improved reliability and resolution KENDA → KENDArtps

23. ISDA 2014, Feb 24 – 28, Munich 23 Power spectrum of ensemble perturbations  Variance at small scales (<100 km) is reduced OPER  Most of the missing variance at small scales developes within 1-2 hours Horizontal wind, model level 30 (~3.1 km), average period (1) +0 h+1 h+3 h

24. ISDA 2014, Feb 24 – 28, Munich 24 Power spectrum of ensemble perturbations  Variance at small scales (<100 km) is reduced OPER  Most of the missing variance at small scales developes within 1-2 hours  Vertical filter: dampening at lower levels exists for more than 3 hours +0 h+1 h+3 h Horizontal wind, model level 40 (~0.8 km), average period (1)

25. ISDA 2014, Feb 24 – 28, Munich 25 How to initialize convective-scale EPS?  Well-known methods for the synoptic global scales, but not clear how to use best for high-resolution limited area models → Downscaling of driving EPS → Ensemble data assimilation at convective-scale COSMO-DE-EPS Δx = 2.8 km - No parametrization of deep convection - 20 ensemble member - 21 hours forecast length - Initialized every 3 hours - Operational since May 2012 → downscaled perturbations of 4 global models + 5 model physics parametrization perturbations → 20 members ~ 1250 x 1150 km