A%20Brief%20Overview%20of%20NCEP%20Short-Range%20Ensemble%20Forecast%20(SREF)%20System

1
A Brief Overview of NCEP Short-Range Ensemble
Forecast (SREF) System
Jun Du, Geoff DiMego and Zoltan Toth
(for the 4th NCEP Ensemble User Workshop, May
13-15, 2008) http//www.emc.ncep.noaa.gov/mmb/SR
EF/SREF.html http//www.nco.ncep.noaa.gov/pmb/nwpr
od/analysis/ Acknowledgement NCEP director, Dr.
Louis Uccellini, and EMC director, Dr. Stephen
Lord, for their always support to the SREF
2
The current SREF configuration

• Entire North America domain (covering U.S.,
  Canada, Mexico, Eastern Pacific and Western
  Atlantic)
• 21 members (with 4 regional models)
• 35-45km depending on models
• Run 4 times per day (03, 09, 15 and 21z)
• 3hrly output to 87hr for each cycle
• 4 AWIPS output domains NorthAmerican (221),
  CONUS (212), Alaska (216) and Eastern
  Pacific/Hawaii (243)

3
Newly added grid221 domain, good for aviation and
hurricane
4
SREF Data

• AWIPS (mean, prob and spgt in grid212, need to be
  in grid221)
• NOMADS (same as above plus individual members,
  need to add Bufr and GEMPAK sounding output?)
• Public ftp site (same as AWIPS plus individual
  members of selected fields, Bufr and GEMPAK
  sounding data will soon be added!)
• NCEP CCS (NCEP internal)
• 2-year archive of all grib outputs (NCEP
  internal). Do we need to archive it at NCDC for
  public?
• Personal help (temporally and limited efforts,
  such as data to RFC and CAPS Spring Pr. Exp.)

5
SREF Products

• Development web (general) http//www.emc.ncep.no
  aa.gov/mmb/SREF/SREF.html
• Operational web (general) http//www.nco.ncep.noa
  a.gov/pmb/nwprod/analysis/
• Aviation products web (by Binbin Zhou) linked at
  the SREF DEV web
• SPC convection (by David Bright)
  http//www.spc.noaa.gov/exper/sref/
• HPC QPF http//www.hpc.ncep.noaa.gov/wwd/impactgr
  aphics/
• WFO (by Rich Gramm) http//eyewall.met.psu.edu/
• Meteograms (by Geoff Manikin) http//wwwt.emc.nce
  p.noaa.gov/mmb/srefmeteograms/sref.html
• Cyclone tracks (by Tim Marchok)
  http//www.emc.ncep.noaa.gov/gmb/tpm/emchurr/tcgen
  /
• Ensemble variance to drive DTRA dispersion model
  http//www.emc.ncep.noaa.gov/mmb/SREF_avia/TEST/we
  b/html/variance.html

6
SREF web (CONUS, AK, HI)
NCEP web
AVN web (CONUS, AK, HI)
SPC?
7
meteograms
Cyclone track
WFO
HPC
8
Comparison to Single Hi-Res Model (SREF21 mean
vs. 12km NAM)
9
Ensemble Mean Performance (SREF vs. GEFS)
10
Which shows that SREF mean outperforms NAM, GFS,
GEFS mean and HPC human forecasts and similar to
the mean of NAM/GFS and the UKMET prior to 60hr
and becomes the best at 72hr. (considering 3hr
earlier in initialization, less perfect ICs used,
courser resolution )
11
However, we are working on the direction of
combining all pieces of information together to
provide users a best possible forecast.
Specifically

• Combining SREF with hi-res single runs (such as
  NAM, GFS, HRW)
• Combining SREF with GEFS

12
SREF21 Spread in Talagrand Distribution
13
Large variance in precipitation forecast within
the SREF
nmm_ctl
arw_ctl
Eta_bmj_ctl
rsm_p1
Eta_kf_ctl
rsm_ctl
14
Reliability of SREF21-based Probabilistic
Forecasts
At 45hr
15
Spread-Skill relation of 500H (upper) and 850RH
(lower) 87hr FCST, 21 SREF members)
500mb H
850mb RH
Forecast error of ensemble mean
Ensemble forecast spread
16
Ens mean worse than hi-res single fcst between
f00-27hr
Spread-skill relation too low between f00-33hr
The major issues now are (1) Addressing 0-1.5d
ensembling strategy including (a) improving ens
techniques, (b) separate ens system focusing
specifically on very short lead time (0-1.5d),
(c) 3hr off, less perfect ICs used, resolution
etc. issues. (2) Improving spread-skill relation
through ensemble technique improvements (IC
pert, physics pert and post-processing).
Relative flat Ranked Histogram reflecting the
underdispersive nature in ens spread is no longer
a major issue but a 2nd one
17
Recent Activities and plans
Systematic error reduction mean error
Systematic Error
Systematic Error Reduction Mean Error
18
SREF-FALL 2007 Implementation Package

• Bias Correction (decaying-average)
• RSM model domain expansion to cover the missing
  part of the AWIPS Alaska 216 grid
• Bufr output from 6 individual WRF members
• 4 aviation products (icing, turbulence, ceiling
  and flight restriction) added for 21 individual
  members
• SREF file directory structure change at CCS to be
  the same as GEFS (NCEP internal)

19
RMSE to measure ensemble-mean forecasts
20
Talagrand Distribution to measure ensemble spread
21
Reliability score to measure probabilistic
forecasts
22
Dynamically Downscaled Ensemble via. Hybrid
Ensembling approach (Du, 2004) Combining the
forecast variance from 40km- SREF and the base
forecast from 4km-HRW wrf runs (44 mems!)
23
nmm0
nmm1
nmm2
40km-WRF members
arw0
arw1
arw2
nmm ctl
obs
Best
arw ctl
4km-HYB members
24
nmm0
nmm1
nmm2
40km-WRF members
arw0
arw1
arw2
nmm ctl
hyb1
hyb2
obs
4km-HYB members
arw ctl
hyb3
hyb4
25
nmm
arw
Eta_bmj
32/45km-SREF members
Eta_kf
rsm
rsm_pert
nmm_ctl
hyb1
hyb2
OBS
4km-HYB members
arw_ctl
hyb3
hyb4
26
Storm-scale High-impact Weather Ensemble Forecast
(HWEF) system focusing on day-1 prediction is
desired because.
27
Case 1 Mar. 14, 2008
40km
12km
4km
OBS
28
Greensburg, KS tornado Day 0600 UTC 5 May 2007,
9 h (SSEF, SPC Spring Program 2007)
CN
P1
radar
N2
29
Dreaming.

• High-impact Weather Ensemble Forecast (HWEF)
  system either Adaptive (relocatable) or
  Fixed-Location (but with multiple focused areas)
  with fine-resolution models.
• Adaptive-to-event system more relevant to severe
  events but technically more difficult and
  psychologically less confident to forecasters due
  to lack of past performance stats. (Beijing
  Olympic system)
• Fixed-Location system more expensive as routine
  production but technically easier and more
  confident in its performance to forecasters. (HRW
  system)
• Similar activities at GSD/ESRL/NOAA (relocatable
  cloud-scale ensemble experiment, VSREF).

30
Plans for 2008-2010

• Upgrade 3 out of the 4 models (nmm, arw and rsm)
  (Fall 2008)
• Replace 4 Eta members with 2 NMM and 2 ARW
  members (Fall 2008)
• Increase horizontal resolution of the 11 members
  from 40/45km to 32km (Fall 2008)
• Dynamically downscaling SREF to 4km
  hybrid-ensemble (need routine run and systematic
  evaluation, how to map with the 5km NDFD grid?)
• Replace Breeding with ET for generating IC
  perturbations (2009)
• Bias correction of precipitation (2009)
• Application to hurricane track forecasting
  (2009?)
• Increase SREF resolution to about 20km (2010)
• High-impact Weather Ensemble Forecast (HWEF)
  system (2010? Computer resource possible?)
• NAEFS-LAM (combining with Canadian SREF, 2010 to
  test?)
• Unifying under ESMF, stochastic physics and
  concurrent run with NAM, GFS, GEFS (beyond 2011)

31
Backup slides
32
SREF Philosophy IC
uncertainty multi-analysis (NDAS and GDAS) and
IC perturbations (breeding) Model physics
uncertainty multi-model (4 models) and
multi-physics schemes (e.g., 6 different
convective schemes) etc. LBC/large-scale flow
forcing uncertainty coupled with
GEFS Lower-boundary forcing uncertainty
perturbing land surface initial states (tested
but not implemented yet) Bias correction
decaying-average statistical approach Downscaling
either dynamical or statistical approaches
(under testing).




33
(AWIPS, Josh)
Image Prob Snow
Image Prob Rain
Image Prob Mix
Image Prob ZR
34
Image Prob gt 1 Snow
(AWIPS, Josh)
Image Prob gt 4 Snow
Image Prob gt 8 Snow
Image Prob gt 12 Snow
35
Ensemble Spread (SREF vs. GEFS)
63hr fcst
36
Probabilistic Forecast (SREF vs. GEFS)
At f45hr
37
Bias corrected variables (32)

• t2m, rh2m, q2m, u10m, v10m, mslp,
• t300, t500, t700, t850,
• z250, z300, z500, z700, z850, z1000
• u250, u300, u500, u700, u850, u1000
• v250, v300, u500, v700, v850, v1000
• rh300, rh500, rh700, rh850

38
SPC Feedback

• SPC
• The SPC has some concerns about the day-to-day
  performance of the weighted decaying average bias
  correction method, especially when applied during
  rapid changes in weather regimes. For example,
  on the synoptic scale during an extended time
  period characterized by a high amplitude blocking
  flow pattern, it is possible that the biases
  computed within this regime may not be
  appropriate when a rapid pattern shift occurs and
  zonal flow then predominates over the area.

39
BC
RAW
BC
RAW
(T2m Bias)
No response to regime change given slow evolving
bias fields and make forecast worse
40
Main Challenge

• Flow dependent bias correction method is needed.
  We are working on it under the framework of
  Bayesian Processor (Prof. Roman Krzystofowicz of
  UVA).

41
Eastern Western QPF bias with WRF_NMM changing
from 5.2km-prod to 4km-para (2007, both explicit
) removal of high-precipitation bias residing
in hi-res forecasts is critical in the Dynamical
Downscaling
Eastern (WRF-NMM), 8/18 to 8/28
Western (WRF-NMM), 8/18 to 8/28
42
Case 2 Apr. 2, 2008
40km
12km
4km
OBS
43
Case 3 Apr. 21, 2008
40km
12km
4km
OBS
OBS
44
Downscaled Variables in the test (18)

• 3h-apcp, t2m, rh2m, q2m, u10m, v10m,
• mslp, cape, cin (9)
• t250, t500, t850, rh850, z500, u250, u850,
• v250, v850 (9)

45
Transiting the Beijing Olympic mesoscale ensemble
system into a relocatable one?
46
Transiting the Operational Hi-Res Window (HRW)
runs into ensemble mode?
Current new Large Domains 4.0 km for WRF-NMM 5.1
km for WRF-ARW Small domain size is unchanged
Before the 2007 change Large Small Domains 5.2
km for WRF-NMM 5.8 km for WRF-ARW
47
Overall Stats for Atlantic Basin 2007
48
Overall stats over Eastern Pacific basin 2007