Hourly RUC Convective Probability Forecasts using Ensembles and

Hourly RUC Convective Probability Forecasts using Ensembles and Radar Assimilation Steve Weygandt Stan Benjamin Forecast Systems Laboratory NOAA

AUTOMATED CONVECTIVE WEATHER GUIDANCE PRESENT • 0 -2 h forecasts from radar extrapolation with growth and decay (nowcasting techniques) • Beyond 2 h guidance from model output helpful FUTURE A seamless convective guidance product utilizing a variety of inputs including nowcasts and model ensemble information to provide guidance to humans and automated decision support systems

Model-based Probability Forecasts for Convective Weather Principle: Convective forecasts at specific model grid points from a single deterministic model run less likely to be correct than averages of model outputs. Procedure: Aggregate model convective information to larger time/space scales (~1 -2 h, 80 -100 km) • Scales should increase with increasing lead time • Scales will decrease as models get better

Ensembles provide technique for aggregating forecast information Types of ensembles • • • Multi-model ensembles Initial/boundary condition ensembles Model physics ensembles Time-lagged model ensembles (2004) Model gridpoint ensembles (2003)

RUC convective precipitation forecast 3 -h conv. precip. (mm) 5 -h fcst valid 19 z 4 Aug 2003

RUC convective probability forecast (2003 -gridpoint ensemble) Threshold > 2 mm/3 h Length Scale = 60 km Box size = 7 GPs 7 pt, 2 mm 5 -h fcst valid 19 z 4 Aug 2003 Prob. of convection within 60 km % 10 20 30 40 50 60 70 80 90

Does probability beat model precip? ----- probability ----- conv precip • Relative “detection” vs. “false-alarm” • “Left and high” curve best detection • Show tradeoff: POD Operating Characteristic (ROC) curves Low prob Low precip 25% High prob 9 pt, 4 mm High precip POFD Sample: 5 -h fcst from 14 z 04 Aug 2003 false detection

Gridpoint Ensembles Adjustable parameters • • Length scale Precipitation Threshold Inherent weaknesses • • Constrained to single model run Non-zero probability can only extend out as far as the characteristic distance More ensemble information better probabilities

Different box sizes and convective precip. thresholds give different probability fields 7 pt, 2 mm 5 pt, 1 mm % 10 20 30 40 50 60 70 80 90 Need to calculate statistical reliability to calibrate probabilities 9 pt, 4 mm 9 pt, 2 mm

Optimal threshold and length scale? 40% 25% 5 -h fcst valid 19 z 4 Aug 2003

RUC Convective Probabilistic Forecast (RCPF) evolution • Automated convective probability forecast • Gridded fields derived from model ensembles • Real-time forecasts started 2003 (RCPFv 2003) • Testing/improvements during 2004 (RCPFv 2004) • 2 -, 4 -, 6 -h forecasts every 2 hours (CCFP guidance) • Verification of forecasts by RTVS • AWC evaluation of product during 2005 • Merge with short-range techniques (NCAR/MIT)

Sample 2003 RUC product RUC Convective Probability Forecast 7 -h fcst valid 21 z 3 Aug 2003 5 pt, 1 mm / 3 h, 40% thresh Threshold probability forecast to get a categorical forecast POD=0. 55 Bias = 1. 4 CSI = 0. 30 Verification display from RTVS

2003 verification of RCPFv 2003 • RCPF most useful • RCPF bias too for initial convective development large all times except evening 6 h Fcst Threshold probability forecast at 40% to get categorical forecast RCPF v 2003 Forecast length Forecast Valid Time GMT EDT Diurnal cycle of convection

Improvements to RCPF for 2004 GOALS (maximize skill) • Reduce large bias (diurnal effects, western differences) • Improve spatial coherency, temporal consistency • Improve robustness • Reduce latency ALGORITHM CHANGES • Increase filter size (9 GP east, 7 GP west) • Time-lagged ensemble (multiple hourly projections from multiple RUC forecast cycles) • Diurnal cycle for precip. thresh. • (maximum daytime, minimum nightime; smaller value in the west) Increase forecast lead time one hour (eg: 6 -h fcst from 13 z valid 19 z available at 1245 z instead of 1345 z)

Diurnal variation of Precipitation Threshold Rate West of 104 deg. longitude, multiply threshold by 0. 6 Higher threshold to reduce coverage Lower threshold to increase coverage Forecast Valid Time GMT EDT Threshold adjusted to optimize the forecast bias - Threshold likely too low at night (bias still too large)

Comparison of RCPFv 2003 and RCPFv 2004 6 h Forecast length Diurnal cycle of convection Forecast Valid Time • • GMT EDT Verification for 26 day period (6 -31 Aug. 2004) RCPFv 2004 fcst is a 1 -h older than RCPFv 2003 RCPFv 2004 has similar CSI, much improved bias

. 20, . 21. 18, . 19. 16, . 17. 14, . 15. 12, . 13. 10, . 11 4 -h 2 -h 6 -h CCFP . 22, . 23 6 -h RCPF v 2004 . 24, . 25 RCPF v 2003 (Verifiation 6 -31 Aug. 2004) Fcst Lead Time CSI by lead-time, time of day 6 -h Forecast Valid Time 4 -h 2 -h Diurnal cycle of convection GMT EDT

2. 5 -2. 75 2. 25 -2. 5 2. 0 -2. 25 1. 75 -2. 0 1. 5 -1. 75 1. 25 -1. 5 1. 0 -1. 25 0. 75 -1. 0 v 2003 2. 75 -3. 0 6 -h CCFP v 2004 (Verifiation 6 -31 Aug. 2004) Fcst Lead Time Bias by lead-time, time of day 6 -h 0. 5 -0. 75 Forecast Valid Time 4 -h 2 -h 6 -h 4 -h 2 -h Diurnal cycle of convection GMT EDT

CSI vs. bias for 2003 vs. 2004 (6 -h forecasts valid 19 z) Low Probabilities 40% Points at 5% intervals 40% High Probabilities • RCPFv 2004 fcst is a 1 -h older than RCPFv 2003 RCPFv 2004 has better CSI for given bias value

Sample RCPFv 2004 product 13 z + 6 h Forecast 19 z RCPF v 2004 At fcst Time. . . 13 z convection verif 25 – 49% 50 – 74% 75 – 100% Verification 19 z NCWD 10 Aug 2004

Sample RCPFv 2004 product 15 z + 6 h Forecast 21 z RCPF v 2004 At fcst Time. . . 15 z convection verif 25 – 49% 50 – 74% 75 – 100% Verification 21 z NCWD 23 July 2004

Interpreting Reliability Plots For all 60% fcsts, event occurs 60% of time (45 deg line) RESOLUTION Strong change in obs freq for given change in fcst probability (vertical line) SHARPNESS OBSERVED frequency (/100) RELIABILITY Tendency forecast probabilities to be near extreme values (0%, 100%) (not hedging) Under forecast y t ili b t ec f er p r ia el Actual reliability Climatology Over forecast FORECAST probability (/100) Tradeoffs between reliability, resolution, sharpness

• Better reliability for 2004 vs. 2003 • Underfcst low prob. , overfcst high prob. • 2004 has many fewer 0% prob. pts that have convection Fractional Coverage • 2004 has more low prob. pts, fewer high prob. pts • 2004 has fewer 0% prob. pts (not shown) FCST fract. areal cover. RELIABILITY OBSERVED frequency (/100) RUC-NCWF 6 -h fcsts valid 19 z 6 -31 Aug. 2004 y t ili b tr c ia el rfe e p Under Climatology Over FORECAST probability (/100) 0. 10 0. 08 0. 06 0. 04 0. 02 0. 00 FORECAST probability (/100)

ACTIVITIES FOR 2005 • Dissemination and evaluation Realtime use and evaluation by AWC Hourly output and update frequency NCAR password protected web-site (model and radar extrapolation) • Ongoing product development Ensemble-based potential echo top information Use of ensemble cumulus closure information Upgrade from 20 -km RUC to 13 -km RUC Use of other RUC fields • Merge RCPF with NCWF 2 (E-NCWF)

Sample RCPF 2005 product 18 z + 6 h Forecast 16 z + 8 h Forecast 2005 RCPF 25 – 49% 50 – 74% 75 – 100% Verification 00 z NCWD 8 Mar 2005 CCFP

Sample Probability/Echo Top Display Probabilities shown with color shading Potential echo top height shown with black Lines (kft) -- Echo top from parcel overshoot level -- Contour echo top height at desired interval (3 kft or 6 kft? )

Grell-Devenyi Cumulus Parameterization • Uses ensemble of closures: - Cape removal - Moisture convergence - Low-level vertical mass flux - Stability equilibrium • Includes multiple values for parameters: - Cloud radius (entrainment) - Detrainment (function of stability) - Precipitation efficiency (function of shear) - Convective inhibition threshold PRESENT: Mean from ensembles fed back to model FUTURE: Optimally weight ensembles closures, Use ensemble information to inprove probabilities

2 hr Nowcast (scale - 60 km) Performance Forecast Closures groups in RUC Grell-Devenyi ensemble cumulus scheme Radar 2100 UTC 10 July, 2002 9 -h fcst valid 21 z 10 Jul 2002

STRENGTHS OF MODEL GUIDANCE • Capturing initial convective development • Long lead-time and early morning forecasts Improvements to the model and assimilation system lead directly to improvements in probability forecasts For RUC model: • Assimilate surface obs throughout PBL • 13 -km horizontal resolution (June 2005) • Radar data assimilation • Full North American coverage (2007)

ISSUES FOR MODEL GUIDANCE • Short-range forecasts (spin-up problem) Poor performance for short-range forecast does not invalidate longer-range forecasts • Propagation of convective systems • Robustness (spurious convection, complete misses) • Model bias issues Differences for parameterized vs. explicit treatments of convection

RUC Radar Data Assimilation Plans Reflectivity: mosaic data • NSSL pre-processing code transferred to NCEP • Integrate mosaic data into RUC cloud analysis • Couple to ensemble cumulus parameterization • Couple to model velocity fields Radial Velocity: level II data • Generalized 3 DVAR solver from lidar OSSE • Use horizontal projection of 3 D radial velocity Outstanding Issues - Data thinning/superobbing - Quality Control (AP, 2 nd trip, unfolding, birds, ) - Optimal uses (clear-air, stratiform precip. , t-storms)

Sample 3 DVAR analysis with radial velocity Cint = 2 m/s 0800 UTC 10 Nov 2004 Dodge City, KS * * * Vr Amarillo, TX Dodge City, KS * * * K = 15 wind Vectors and speed Cint = 1 m/s * Vr Amarillo, TX 500 mb Height/Vorticity Analysis WITH radial velocity * Analysis difference (WITH radial velocity minus without)

Thoughts and questions Predictability very limited for small-scale convective precipitation features • Smoothing improves many scores • Smoothing alters spectra, probability information Many “radar” approaches applicable to model forecast precipitation fields • Probabilities from spatial variability of model precip. • Model depicts “displacement”, and “temporal evolution” • Apply “tracking” algorithms to model precipitation fields? Many opportunities for blending model- and radar-based techniques • Need extensive comparison to find “break even” points • Assess ability of radar and model for different tasks • Merge radar structure with model favored regions?

CONVECTIVE STORM TYPE Squall-line 30% • Discernible from probability shape 50% Not as clear for other shapes 70% • Scattered storms (high likelihood, 20% coverage) • MCS (20% likelihood, significant coverage) Storm-type affects correlation of adjacent probabilities, cumulative probability for flight track 30%

How is the RCPF created? 1. Gridpoint ensemble (for each model GP) - Fraction of 20 -km model gridpoints within 9 x 9 box with 1 -h convective precipitation exceeding threshold (use 7 x 7 km box west of 104 deg. Longitude) - Diurnal variation to 1 -h convective precipitation threshold (smaller value for threshold west of 104 deg. longitude) 2. Time-lagged ensemble - Use up to six forecasts bracketing valid time - 9 -h RUC forecast every hour with hourly output - 2 -h latency to RUC model forecast output 4 -h RCPF inputs M 0+4 M 1+5 M 2+6 M 0+5 M 1+6 M 2+7 6 -h RCPF inputs M 0+6 M 1+7 M 0+7 M 1+8 8 -h RCPF inputs M 0+8 M 1+9 M# = # hours M 0+9 back to model initial time

14 z 15 z 16 z 17 z 18 z 19 z 20 z 21 z 22 z 23 z Available Initial Time 12 z 13 z 14 z 15 z 16 z 17 z 18 z 19 z 20 z 21 z RCPF has 2 h latency Time-lagged ensemble inputs RUC model HHz = model intial time forecasts F = forecast length (h) (HHz+F) 15 z RCPF (17 z CCFP) 15+9, 10 15+2, 3 15+3, 4 15+4, 5 15+5, 6 15+6, 7 15+7, 8 15+8, 9 14+5, 6 14+6, 7 14+7, 8 14+8, 9 14+9, 10 14+10, 11 14+3, 4 14+4, 5 13+5, 6 13+6, 7 13+7, 8 13+8, 9 13+9, 10 13+10, 11 13+11, 12 2 2 14 z RCPF (16 z CCFP) 3 4 4 5 6 6 7 8 9 8 14 z+2, 3 14 z+4, 5 14 z+6, 7 14 z+8, 9 13 z+3, 4 13 z+5, 6 13 z+7, 8 13 z+9 12 z+4, 5 12 z+6, 7 12 z 13 z 14 z 15 z 16 z 17 z 18 z 19 z 20 z 21 z 22 z 23 z 00 z Forecast Valid Time (UTC)