Toward improved initial conditions for NCAR s real-time convection-allowing ensemble. Ryan Sobash, Glen Romine, Craig Schwartz, and Kate Fossell

Transcription

1 Toward improved initial conditions for NCAR s real-time convection-allowing ensemble Ryan Sobash, Glen Romine, Craig Schwartz, and Kate Fossell

2 Storm-scale ensemble design Can an EnKF be used to initialize a convectionallowing ensemble to produce skillful next-day predictions of convection? Different than other storm-scale ensemble designs Ensemble analysis (DART), continuously cycled (WRF) Large nest domain to push lateral boundary condition uncertainty away from area of forecast interest Single WRF physics suite with explicit representation of convection (identify and mitigate model errors)

3 NCAR Ensemble-based cycled analysis with DART X a = X f +K[y 0 HX f ] Ensemble background (X f ) WRF Member 1 WRF Member 2 WRF Member 3 observations (K) DART filter (EAKF) Model estimate observations (HX f ) Ensemble analysis (X a ) WRF Member 1 WRF Member 2 WRF Member 3 (y 0 ) WRF model integration (N=80) WRF ensemble forecasts

4 NCAR ensemble domains 15 km grid spacing GFS + perturbations Analysis domain (N = 80) GFS + perturbations 3 km grid spacing GFS + perturbations GFS + perturbations Lower boundary: free forecast land surface, fixed sea state during integration Forecast domains (downscaled; N = 10)

5 Observations used in the analysis system Routine observations from a variety of sources, not uniformly distributed in type or time. Simplified observation suite compared to operational prediction centers. No radiances are assimilated.

6 Accessible ensemble forecast information

7 Ensemble forecast verification Forecast reliability (2013 system): hour precipitation Ensemble Deterministic

8 Ensemble forecast verification Forecast reliability (2015 system): hour precipitation

9 Ensemble forecast verification

10 Ensemble forecast verification Forecast reliability (2013 system): Severe weather events Observed Relative Frequency! (a)! # points! (b)! Forecast Probability! Sobash et al (WAF)

11 Improving forecast reliability How can we improve reliability of forecasts? 1. Post-processing methods Need large sample of past cases for rare events 2. Improvements to analysis system Examine impact of inflation settings Test alternative inflation algorithms Do different spread techniques produce different patterns of uncertainty?

12 Real-time analysis system inflation exps Type of Inflation Anderson et al. (2009) Anderson et al. (2009) Anderson et al. (2009) Multiplicative inflation experiments Prior or Posterior Inflation Prior (SD=0.7; damp=0.9) Posterior (SD=0.7; damp=0.9) Prior+Posterior (SD=0.7; damp=0.9) Whitaker and Hamill RTPS Posterior (α = 1.24 and 1.0) Retrospective testing with cycled EnKF system from 27 April 31 May 2015

13 Real-time analysis system inflation exps After four weeks of cycling 850 hpa Temperature mean and spread (m) PRIOR

14 Real-time analysis system inflation exps After four weeks of cycling 850 hpa Temperature mean and spread (m) POST

15 Real-time analysis system inflation exps After four weeks of cycling 850 hpa Temperature mean and spread (m) PRIOR+POST

16 Real-time analysis system inflation exps After four weeks of cycling 850 hpa Temperature mean and spread (m) WHIT

17 Real-time analysis system inflation exps RMSE Temp (K) U-wind (m/s) Prior RMSE/Spread Compared to radiosondes Spread

18 Looking ahead: Improving EnKF ICs Lightning flash rate Ongoing work to improve analysis system (2016) Inflation experiments: better tuning for inflation settings Assess impact on hour ensemble forecast skill. EnKF cloud analysis (dx ~ 3 km) on the current prediction grid (2017 and beyond) 1-12 h prediction window, multiple times per day More frequent analyses (at least hourly) Utilize radar, satellite, lightning, mesonet observations Awaiting next supercomputer: Cheyenne

19 Analysis system updates for Spring 2016 Analysis system ensemble size increased from 50 to 80 members. Introduction of spread restoration to be discussed Began assimilating GPS RO datasets

20 Real-time analysis mean innovations August 2015 mean analysis innovations (O-F) for 00 UTC Lowest model level temperature Lowest model level water vapor Classic MYJ bias across plains and southeast US, opposite of this in other areas.

21 Real-time analysis mean innovations December 2015 mean analysis innovations (O-F) for 00 UTC Lowest model level temperature Lowest model level water vapor Slight cool bias in December but regional variability, dry along Gulf Coast

22 Ensemble forecast verification Fractions skill score: Precipitation Ensemble Deterministic

23 Ensemble forecast verification Fractions skill score: Severe weather events Ensemble Fractions Skill Score! Deterministic σ (km)! Sobash et al (WAF)