430 likes | 566 Vues
Real Time Mesoscale Analysis. RTMA Temperature 1500 UTC 14 March 2008. John Horel Department of Meteorology University of Utah john.horel@utah.edu. Acknowledgements Dan Tyndall (Univ . of Utah) Dave Myrick (WRH/SSD ) Manuel Pondeca (NCEP/EMC) References
E N D
Real Time Mesoscale Analysis RTMA Temperature 1500 UTC 14 March 2008 John Horel Department of Meteorology University of Utah john.horel@utah.edu
Acknowledgements • Dan Tyndall (Univ. of Utah) • Dave Myrick (WRH/SSD) • Manuel Pondeca (NCEP/EMC) • References • Benjamin, S., J. M. Brown, G. Manikin, and G. Mann, 2007: The RTMA background – hourly downscaling of RUC data to 5-km detail. Preprints, 22nd Conf. on WAF/18th Conf. on NWP, Park City, UT, Amer. Meteor. Soc., 4A.6. • De Pondeca, M., and Coauthors, 2007: The status of the Real Time Mesoscale Analysis at NCEP. Preprints, 22nd Conf. on WAF/18th Conf. on NWP, Park City, UT, Amer. Meteor. Soc., 4A.5. • Horel, J., and B. Colman, 2005: Real-time and retrospective mesoscale objective analyses. Bull. Amer. Meteor. Soc., 86, 1477-1480. • Tyndall, D., 2008: Sensitivity of Surface Temperature Analyses to Specification of Background and Observation Error Covariances. M.S. Thesis. U/Utah
An Analysis of Record • Forecasters have needs for higher resolution analyses than currently available • Localized weather forecasting • Gridded forecast verification • Climatological applications • AOR program established in 2004 • Three phases • Real Time Mesoscale Analysis • Delayed analysis: Phase II • Retrospective reanalysis: Phase III
Real-Time Mesoscale Analysis (RTMA) • Fast-track, proof-of-concept intended to: • Enhance existing analysis capabilities at the NWS and generate near real-time hourly analyses of surface observations on domains matching the NDFD grids. • Background errors can be defined using characteristics of background fields (terrain, potential temperature, wind, etc.) • Provide estimates of analysis uncertainty • Developed at NCEP, ESRL, and NESDIS • Implemented in August 2006 for CONUS (and southernmost Canada) & recently for Alaska, Guam, Puerto Rico • Analyzed parameters: 2-m T, 2-m q, 2-m Td, sfc pressure, 10-m winds, precipitation, and effective cloud amount • 5 km resolution for CONUS with plans for 2.5 km resolution
The Real-Time Mesoscale Analysis • Several layers of quality control for surface observations • Two dimensional variational surface analysis (2D-Var) using recursive filters • Utilizes NCEP’s Gridpoint Statistical Interpolation software (GSI) • Uses various surface observations and satellite winds • METAR, PUBLIC, RAWS, other mesonets • SSM/I and QuikSCAT satellite winds • Analysis became available in 2006, and is being evaluated by WFOs and several universities
The actual ABCs… • The RTMA analysis equation looks like: • Covariances are error correlation measures between all pairs of gridpoints • Background error covariance matrix can be extremely large • 2,900 GB memory requirement for continental scale • Recursive filters significantly reduce this demand
CVOC & CVOI CVOD RTMA Vancouver Area
Background DownscalingBenjamin et al. (2007) • CONUS RTMA background = 1-h forecast from the NCEP-operational 13-km RUC downscaled to the 5-km NDFD terrain • Horizontal - bilinear interpolation • Vertical interpolation – varies by variable, for temperature it is based on near-surface stability and moisture from the RUC native data used to adjust to the RTMA 5-km terrain • If RTMA terrain lower than RUC, then local RUC lapse rate used (between dry adiabatic and isothermal) • If RTMA terrain higher than RUC, interpolated between native RUC vertical levels, but shallow, surface-based inversions maintained • Coastline sharpening
Surface Data Issues • Real-time RTMA analysis begins ~30 min past the hour • Getting the data in time is a challenge • RTMA uses obs taken (+/-12 min from top of hour) • RAWS and Quickscat winds (-30 min to +12 min) • Many remote obs still don’t make it in time, as well as some obswith longer latencies (Snotel = every 3 hrs) Observation Network MADIS NCEP
Subjective Evaluation of the RTMA Temperature • Development of web based graphics for many analyses over selected subdomains • Wasatch Valley, UT • Puget Sound, WA • Norman, OK • Shenandoah Valley, VA • Analysis problems based on subjective evaluation • Non-physical treatment of temperature inversions • Smoothed features • Bull's-eye features around observations • Overfitting issues • Case study • 0900 UTC 22 October 2007 • Shenandoah Valley, VA
Case Study Orientation RTMA 0900 UTC 22 October 2007
RTMA Topography vs. Actual Topography 0 100 200 300 400 500 600 700 800 900 1000
1200 UTC 22 October 2007 KIAD Sterling, VA Atmospheric Sounding
RUC 1-hr Forecast Valid 0900 UTC 22 October 2007 RUC 1-hr Forecast
RTMA 0900 UTC 22 October 2007 RTMA Temperature Analysis
RTMA Temperature Analysis Increments RTMA Temperature Analysis Increments 0900 UTC 22 October 2007
Observations • Surface observations: METAR, RAWS, PUBLIC, OTHER • Observations must fall in ±12 min time window (−30/+12 min for RAWS) • Observations undergo quality control by MADIS and internal checks by RTMA • Approximately 11,000 observations for almost 900,000 gridpoints for CONUS
METAR 16/59/1,744 OTHER 10/75/1,961 Observation Density PUBLIC 215/575/6,486 RAWS 3/11/1,301 Observation Density
Observation and Background Error Variances • Ratio of σo2/σb2estimated by accumulating statistics over a time period for entire CONUS • 8 May 2008 – 7 June 2008 • RTMA used σo2/σb2 = 1 for METAR observations and σo2/σb2 = 1.2 for mesonet observations • Statistics computed in our research suggest σo2/σb2 between 2 and 3 • Background should be “trusted” more than observations
Estimation of Observation and Background Error Covariances • Temperature errors at two gridpoints may be correlated with each other • Error covariances specify over what distance an observation increment should be “spread” • RTMA used decorrelation lengths of: • Horizontal (R): 40 km • Vertical (Z): 100 m
Error Correlation- KOKV R = 40 km, Z = 100 m Error Correlation Example – KOKV
Error Correlation- KOKV R = 80 km, Z = 200 m Error Correlation Example – KOKV
Local Surface Analysis • RTMA experiments run on NCEP’s Haze supercomputer but limited computer time available • Development of a local surface analysis (LSA) • Same background field • Same observation dataset, but without internal quality control • Similar 2D-Var method, but doesn’t use recursive filters • Smaller domain • Control run uses same characteristics of RTMA (decorrelation length scales; observation to background error ratio) at time of study
LSA R = 40 km, Z = 100 m, σo2/σb2 = 1 LSA Temperature Analysis
RTMA 0900 UTC 22 October 2007 RTMA Temperature Analysis
Data Denial Experiments • Evaluation of analyses done by splitting up all observations randomly into 10 groups • Two error measures: • Root-mean-square error calculated using the withheld observations only vs. using the observations assimilated into the analysis • To avoid overfitting, want RMSE at withheld locations to be similar to RMSE at locations used in the analysis • Root-mean-square sensitivity computed at all gridpoints • Want analyses to be less sensitive to withheld observations
Group 5 R = 40 km, Z = 100 m, σo2/σb2 = 1 Withholding Observations
Withheld Data Groups 1-5 R = 40 km, Z = 100 m, σo2/σb2 = 1 -1 -0.5 0 0.5 1 2.5 4.5 6.5 8.5 10.5 12.5 14.5 16.5
Withheld Data Groups 6-10 R = 40 km, Z = 100 m, σo2/σb2 = 1 -1 -0.5 0 0.5 1 2.5 4.5 6.5 8.5 10.5 12.5 14.5 16.5
Group 1 RTMA RTMA Temperature Analysis Sensitivity
LSA RMSE and Sensitivity Compared to RTMA Suggests overfitting Overly sensitive to withheld observations
How good does the RTMA have to be? • Is the downscaled RUC background field good enough? • RUC RMSE approximately 2°C to all observations in CONUS • How much does a 2D-Var analysis improve RMSE? • 0.1-0.2°C based on LSA results
Verifying Forecasts with Analyses • Motivating Question: Is there a way we can define what constitutes a “good enough” forecast? • Problem: Concerns about grid-based verification • Reality: Forecasters need feedback for entire forecast grids not just selected key observation locations
Forecaster Concerns • Quality of the verifying analysis • Penalty for adding mesoscale detail to grids in areas unresolved by analysis • Bad (mesonet) observations influencing analysis • Analysis in remote areas – driven mostly by the background model
Using RTMA Analyses and RTMA Uncertainty Estimates for Forecast Verification • RTMA uncertainty estimate grids are experimental products under development for temperature, moisture, wind • Goal: • Higher uncertainty in data sparse areas or areas with larger representativeness errors • Lower uncertainty in data dense areas or areas with smaller representativeness errors • Basic premise • Forecasters not penalized as much where the analysis is more uncertain
Temperature (oC) Forecast Verification Example 2 6 8 1 7 Valley w many obs 1 10 Mountains w unrepobs 8 2 12 3 9 4 Forecast RTMA RTMA Uncertainty
Temperature (oC) Forecast Verification Example 2 6 8 1 7 Valley w many obs 1 10 Mountains w unrepobs 8 2 12 3 9 4 Forecast RTMA RTMA Uncertainty No color= Differences between the forecast and analysis are less than analysis uncertainty Red = abs(RTMA – Forecast) > Uncertainty
Summary • Improving current analyses such as RTMA requires improving observations, background fields, and analysis techniques • Increase number of high-quality observations available to the analysis • Improve background forecast/analysis from which the analyses begin • Adjust assumptions regarding how background errors are related from one location to another • Future approaches • Treat analyses like forecasts: best solutions are ensemble ones rather than deterministic ones • Depend on assimilation system to define error characteristics of modeling system including errors of the background fields • Improve forward operators that translate how background values correspond to observations