Data Assimilation and Predictability Studies for Improving TC Intensity Forecasts PI: Takemasa Miyoshi University of Maryland, College Park Co-PIs: E. Kalnay and K. Ide (UMD) and C. Bishop (NRL) Co-Is: T. Enomoto and N. Komori (Japan), S.-C. Yang (Taiwan), H. Li (China) Collaborators: T. Nakazawa (WMO), P. Black (NRL) Project Researcher: M. Kunii (UMD) February 2011, NOPP TC Topic Review, Miami
Adaptive inflation/localization Ensemble sensitivity methods Lagrangian data assimilation Better use of observations Project Overview CFES-LETKF using the Earth Simulator Coupled Atmos-Ocean error covariance Large-scale environment DA Method LETKF Local Ensemble Transform Kalman Filter (Hunt et al. 2007) WRF-LETKF Cloud-resolving data assimilation Mesoscale
Achievements Studies on methods –Adaptive inflation method was developed. –Impact of resolution degradation was investigated. CFES-LETKF –AFES-LETKF experiments were performed. WRF-LETKF –WRF-LETKF system was developed and tested. –Impact of SST perturbations was examined. –Observation impacts were computed for T-PARC dropsonde observations.
Deliverables Peer-reviewed publications –3 accepted Miyoshi and Kunii (2011, Pure Appl. Geophys.) on the WRF-LETKF Miyoshi (2011, MWR) on the adaptive inflation method Miyoshi et al. (2010, WAF) on the resolution degradation of the initial condition –1 submitted Miyoshi et al. on the observation error correlations Conference presentations –9 orals (2 invited), 2 posters
ADAPTIVE INFLATION Studies on methods:
Variance underestimation T=t0T=t1 An initial condition with errors FCST Ens. mean P Forecast ensemble tends to be under-dispersive.
Covariance inflation T=t0T=t1 An initial condition with errors FCST Ens. mean P (1+a)P Covariance inflation inflates the underestimated variance.
Adaptive inflation Posterior PriorObs Li et al. (2009) applied the Gaussian assumption. PosteriorPriorObs Non-Gaussianity is very weak.
Localization of inflation estimates Each grid point is treated independently. Local Ensemble Transform Kalman Filter (LETKF, Hunt et al. 2007) Apply the maximum likelihood estimate at each grid point. PosteriorPriorObs
Test with a T30/L7 AGCM
Adaptive inflation estimates
Impact of adaptive inflation ~20% improvements of analysis RMSE Better match between RMSE and spread
AFES-LETKF CFES-LETKF:
Experimental settings AFES-LETKF settings ResolutionT119/L48 Ensemble size63+1 Covariance inflationFixed 10% Covariance localization400 km, 0.4 ln p 4D-LETKF
Surface SondesAircraft Ships & buoys QuikScatAMV
Analysis of TY Sinlaku (2008) Best track AFES-LETKF
Analysis and ensemble spread
REAL OBSERVATIONS WRF-LETKF:
Experimental settings LETKF settings WRF settings Ensemble size20 Lateral boundary conditionsUnperturbed Covariance inflationAdaptive (Miyoshi 2010) Fixed 20% (smaller above level 20) Covariance localization400 km, 0.4 ln p Analyzed variablesu, v, w, T, ph, qv, qc, qr Domain size136(137) x 108(109) x 39(40) Horizontal grid spacing~ 60 km WRF versionWRF-ARW 3.2
6-hr Forecast at 12Z 12 Sep NCEPGSMAP NOOBSLETKF after 9 days cycle
Adaptive inflation Lower troposphereMiddle troposphere ● ADPSFC ▲ SFCSHP □ ADPUPA ◇ AIRCFT ○ PROFLR × SATWND ■ VADWND △ SPSSMI + QKSWND Adaptive inflation accounts for imperfections such as model errors and limited ensemble size. Large adaptive inflation > 100 % (2.0) appears occasionally and is appropriate in the limited regions.
Ensemble spread (T500) ADAPTIVEFIXED20% NOOBS Too small (large) over the densely (sparsely) observed areas Unperturbed BC is limited near the boundaries
6-hr forecast vs. radiosondes Adaptive inflation performs well.
Computational time per 6-h cycle Computing environment: Number of nodes: 7 (Linux servers) CPUs per node: 2 x Quad-Core AMD Opteron 2382 (2.6 GHz) Total number of cores: 7 x 8 = 56 # mem.cores per fcst. mem. mem. per node total cores for fcst. cores for LETKF * * * * 1642 * +1: forecast from ensemble mean analysis
Impact of SST perturbations Consistently better with SST perturbations.
OBSERVATION IMPACT WRF-LETKF:
Introduction Traditionally, observation impact has been estimated by carrying out the data denial experiment (expensive). Methods without data denial experiment: –Adjoint-based method (Langland and Baker 2004) –Ensemble-based method (Liu and Kalnay 2008) In this study –Estimate impact of real observational data with the WRF-LETKF system using the ensemble-based method.
Experimental settings TY0813 (SINLAKU) Track MSLP OBSIMP settings Verification time6 h, 12 h NormKE Targeted regionNONE, 5 × 5° around the TC
Observation impact for each type Forecast error reduction (J/kg, KE) Observation count Observation impact for each observation type (except for satellite radiances) (9/8 12 UTC – 9/12 12UTC, NW Pacific)
Observation Impact 00UTC Sep Forecast error reduction (J/kg, KE) SONDE AMVSPSSMI
Targeted Norm No targeting Targeting (5 × 5° around TC ) 00UTC Sep Forecast error reduction (J/kg, KE) in the targeted area ~1460 km Radius of influence
Impact of DOTSTAR dropsondes 6 h fcst error reduction12 h fcst error reduction J/kg
Denying negative impact data 6 h fcst error reduction12 h fcst error reduction J/kg
Denying negative impact data improves forecast! 6hr fcst error is reduced Estimated observation impact 6-h track forecast is actually improved!! 1.5-day forecasts
Denying negative impact data improves forecast! 12hr fcst error is reduced Estimated observation impact 12-h track forecast is actually improved!! 1.5-day forecasts
Impact of NRL P-3 dropsondes
Impact of WC-130J dropsondes
Impact of DLR Falcon dropsondes
FUTURE PLANS
Plans More focusing on TC intensity forecasting –Investigating air-sea coupled covariance around TC –Higher-resolution experiments to resolve TC structures –Predictability studies by ensemble prediction –More TCS-08 case studies More analysis of the impact of aircraft observations –Impacts of lower-level and upper-level flights at each stage Use of rapid-scan cloud images Direct assimilation of best-track data –Possibly, further ITOP-10 case studies Cases with joint DOTSTAR and WC-130J missions
The LETKF code is available at:
RESOLUTION DEGRADATION Studies on methods:
Experience from operations at JMA T959 operational forecasts T319 forecasts from T959 analysis (operational Typhoon EPS) T319 forecasts from T319 analysis Simply eliminating higher wavenumber components has negative impact on TC track forecasts.
Methods to initialize T319 1.Smooth spectral truncation 2.One-time low-resolution DA Operational Typhoon EPS Test1: Smooth truncation Test2: One-time DA
Results of Typhoon Nuri (2008) Best track observation T959 operational forecast T319 forecasts (operational EPS) T319 forecasts from T319 analysis Test1: Smooth truncation Test2: One-time DA
Near-center TC structure Bad T319 forecast Better T319 forecast Operational T959 Conclusion: It is important to consider the model’s resolving capability in data assimilation T319 analysis