Prediction Thrust Activities in CASA Briefing for WNI
Prediction Thrust Activities in CASA Briefing for WNI Dr. Ming Xue Director of CAPS Center for Analysis and Prediction of Storms August 2, 2006
High-resolution simulations and tornadoes, tornadic thunderstorms and downbursts
25 -m tornado simulation using 2048 1 GHz alpha processors. 30 -min simulation took 18 hours
25 m resolution simulation of a tornado within a supercell thunderstorm Numerical Simulation by Ming Xue School of Meteorology University of Oklahoma Movie of Cloud Water Field 25 m, 7. 5 x 7. 5 km domain, 30 minutes Movie by Greg Foss Pittsburgh Supercomputing Center
See 2 movies
Variational Analysis of Over-sampled Dual-Doppler Radial Velocity Data and Application to the Analysis of Tornado Circulations Ming Xue 1, Shun Liu 1 and Tian-you Yu 2 1 Center for Analysis of Prediction of Storms University of Oklahoma Based on Xue, Liu and Yu (JTech 2006, Accepted)
Motivation • CASA radars will have 100 m radial resolution but 2 degree beam width therefore coarse azimuthal resolutions • The azimuthal resolution is most important for capturing wind shear associated with tornado circulations • Emulator-based studies have shown that the broad beams of CASA radars present the largest challenge with tornado detection/identification • More detailed flow structure can be recovered from over-sampled dual-Doppler wind measurements • Variational method combined with a realistic observation operator/radar emulator can retrieve subbeam structures smoothed out by the broad beam-
Sample Volume Illustration of the simulation of radial velocity data from a gridded wind field. Similar to that of Wood and Brown (1997), except for oversampling.
Principle • In the following 1 -D illustration, when the data samples are the same in number as the number of grid points, the grid point values can be determined exactly. • Equation is Ax = y, where y is observation vector, x is the state vector on the grid, and A contains the coefficients of discrete weighting function. We seek x from y. • Our problem is more difficult because u and v are not directly observed (Vr is), and observations may be insufficient to determine u and v. • Variational method is most suitable for such a problem – under-determinedness can be avoided by including a background and/or spatial smoothing – realized via recursive filter • Dual-Doppler wind observations are assumed an
Radial Velocity Data Emulation Vortex is assumed 2 D Wood and Brown (1997)
Cost function of variational analysis • Bu and Bv are modeled using recursive filter. • Spatial de-correlation scale was 5 grid intervals for idealized case and 10 grid interval for real case • Background wind is assumed zero.
Tests using Simulated Data • Simulated radar observations for the 25 m tornado simulation of Xue • Analysis domain is 9 x 9 km. Grid size is 361 x 361 with 25 m grid spacing. • Range resolution is 100 m. • Experiments examine the impact of the following on analysis: – azimuthal increment of over-sampling, – distance of radars from tornado, – effective beamwidth
Simulated radial velocity fields 2 o Radar is located at (0, 0) km, or 15 km south of the 9 km × 9 km analysis domain center with azimuthal increments of (a) 2 o, (b) 1 o and (c) 0. 125 o (c) 1 o 0. 125 o Over-sampling alone, without wind retrieval, is not much of a help beyond a factor of two over-sampling
Impact of azimuthal increments of samplin u “truth” & analyses from radial velocity data sampled at azimuthal increments of 0. 125 o and 2 o v wind truth O/S 0. 125 o Beamwidth = 2 o Distance of radars = 15 km N/O 2 o
Correlation coefficients (CC) for different azimuthal increments when the radars are located 15 km from the center of the analysis domain. Azimuthal increment 0. 125 0. 5 1. 0 1. 5 2. 0 CC 0. 90 0. 82 0. 78 0. 72 0. 68
Impacts of radar distance from the tornado Correlation coefficients of the analyses when the radars are located at different distances from the center of analysis domain, for azimuthal increments of 0. 125º and 2º. Distance 12. 0 15. 0 18. 0 21. 0 24. 0 Azimuthal increment 0. 125º 0. 92 0. 90 0. 83 0. 82 0. 76 Azimuthal increment 2º 0. 71 0. 69 0. 66 0. 65 0. 56
Impacts of effective beamwidth 1 o 2 o Increment = 0. 125 o , beamwidth = 1 o or 2 o , distance=15 km CC = 0. 95 for 1 o, 5% > CC of 2 o beamwidth case
Tests using over-sampled KOUN data and regular KTLX data • Level-I data for May 8, 2003 OKC tornado from KOUN reprocessed to over-sample at 1/8°increments. 64 pulses are used in each sample. • Data from KTLX at regular 1°sample intervals are used together in the variational dual-Doppler analysis
Radar radial velocity observations from KTLX and KOU for May 8 th, 2003 OKC tornado case KTL X KOUN 1. 0 o azimuthal increment KOUN 0. 125 o azimuthal increment
Analyzed wind fields with 8 times oversampling and without oversampling for KOUN 8 times oversampling with KOUN 35 60 no oversampling u v V
Summary • Significantly more details of the flow can be recovered through variational analysis from over-sampled dual-Doppler winds, which is important for characterizing tornado circulation and tornado detection • For simulated data, when the azimuthal increments (~0. 1 degree @ 15 km range & 1 o effective beamwidth) are comparable to the ‘truth’ resolution (25 m), the tornado circulation can be analyzed rather accurately (CC=0. 95). • The analysis is still smoother than the ‘truth’, due to the need for the help of background error (spatial) covariances (because of the under-determinedness issue at least at the far ranges), realized through spatial smoothing • Tests with a real tornado case indicates that reliable oversampled data can be obtained through re-processing Level-I data, for the WSR-88 D radars, and the wind analysis using ‘over -sampled’ data from one radar alone is shown to improve dual. Doppler wind analysis for tornado.
Closed-loop peak echo tracking STSM nowcaster operational 4 node mosaic’d data operational 6/06 Implement differential phase attenuation: 7/1/06 Closed-loop peak echo tracking: 7/8/06 GMAP clutter removal operational: 8/15/06 Network-based attenuation R&D using IP 1 data: 7/1/06 – 9/1/06 Implement & test network-based attenuation: 9/1/06 – 12/1/06 Network-based attenuation operational: 5/07 PTDM clutter R&D using IP 1 data: 7/1/06 – 8/31/06 Implement PTDM in IP 1: 9/1/06 – 12/1/06 Implement 4 node mosaic using REORDER 7/06 – 7/06 4 node mosaic’d data operational 9/06 Implement STSM nowcaster in IP 1: 8/06 – 10/06 STSM nowcaster R&D using IP 1: 11/06 – 5/07 STSM nowcaster operational: 5/07 2 km forecasts using IP 1 and ADAS: 5/07 1 km forecasts using IP 1 and 3 DVAR: 5/08 1 km forecasts using 3 DVAR 2 km forecasts using ADAS 6/07 Project ends 6/08
Nowcasting • STSM nowcaster operational (2/2007) • WDSS-II Storm Cell Identification and Tracking (SCIT) • NWP model-based very short-range forecasting
Nowcasting v. s. NWP should be here! (Jim Wilson 2006, US-Korea Workshop) • The quoted NWP models are not initialized with radar and/or other high-res data and/or their resolution is too low initialized NWP models should out-perform • Properly extrapolation/statistical nowcasting models from time zero • Simple, fast nowcasting systems are useful for very short ranges because they are fast, not because they are intrinsically better
Real Forecast Examples: May 8 th, 2003 OKC tornado 2210 -2238 UTC 30 km long path F 4 Tornado #1 2200 UTC 2204 -2210 UTC
ARPS 1 -km-grid forecast using 3 DVAR + Cloud Analysis cycles over 1 hour Reflectivity at 1. 45º elevation 30 -min forecast 40 -min forecast
Observed v. s. Predicted Z and Vr at 1. 45° Observation 1 km Forecast Reflectivity Radial velocity From 2140 to 2240 UTC every 5 -min
Prediction using 100 m resolution grid sfc winds ck a r t do na pert. pressure tor. s b o (over 22 minutes)
NWP v. s. Non-dynamic Nowcasting • State-of-the-art NWP (model + DA) systems are capable of predicting real tornadoes now! • There is real hope for ‘warn on prediction’, even for tornadoes!
Time line of planned real time implementations of data assimilation (DA), adaptive observation (AO) and prediction systems for CASA DA and Prediction System Features and Functionalities Start of realtime operation Nowcasting Storm Tracking and Nowcasting System; 1 – 60 minute forecast, 1 -5 minute forecasts used in MC&C for closed loop control Spring 2007 1 st generation DA & NWP ADAS/3 DVAR and Cloud Analysis initialization of ARPS, 5 -10 minute update cycles; 30 minute to 6 hour forecast Spring 2007, ADAS-based system, ~2 km resolution; Spring 2008, 3 DVAR-based system, ~ 1 km resolution 2 st generation DA & NWP En. KF DA with improved NWP model; 30 -second analysis cycles; 30 minute to 6 hour forecast Spring 2009 3 st generation DA, AO & NWP Ensemble-based DA, AO and prediction; fully functional ETKF-based adaptive sampling, En. KF data assimilation and ensemble prediction system; scanning strategy optimized forecast over 30 minute to 6 hour is decided before the scanning time and integrated into the MC&C. Assimilating clear air data and polarimetric parameters. Spring 2010 -2011, within IP 5.
Planned domain for 2007 CAPS spring forecast project for LEAD and CASA • Resolution 2 – 4 km • Three 2 -km forecasts, one 30 hours and two 9 hours long • 8 -member 4 km ensembles • Assimilating WSR 88 D and CASA radar data in the two 9 -h 2 km forecasts • Requires 8, 400 CPU hours/day • Using ARPS and WRF • Will need about 2000 CPUs concurrently to do all of the above.
Main Areas Research of Analysis and Prediction Thrust (mostly at OU) • • Radar emulation Tornado detection and anticipation algorithms Real time analysis En. KF and 3 DVAR data assimilation Storm-scale NWP Dual-pol data assimilation Adaptive observing system development Optimal scanning strategies for detection and NWP • Tornado dynamics and phenomenological studies using CASA radar data
Hazardous Weather Detection • Adaptation and tuning of WDSS-II WSR-88 D-based severe weather detection algorithms to work with CASA Networked radar data (MDA/TDA, LLSD and wind analysis) • Observing platform-independent detection algorithms based on highresolution analysis/assimilation data sets, that make full use of all available data (Fritchie et al. 2005; Xue et al. 2006) • Wavelet analysis-based tornado detection algorithm (TDA, Liu et al. 2006) • TDA based on spectrum/time series data (Yu et al. 2003) • Tornado characterization and detection via optimal fitting to data of loworder tornado vortex models (Potvin et al. 2005) • Tornado anticipation algorithms via data mining/pattern recognition (Rosendahl, Droegemeier and Mc. Govern) • Hydrometeor classification, wind analysis and rain rate estimation (Brenda Dolan)
- Slides: 33