New 16km NCEP Short-Range Ensemble Forecast (SREF)
system: what we have and what we need?
(SREF.v6.0.0, implementation date: Aug. 21. 2012)
Jun Du, Geoff DiMego, Binbin Zhou, Dusan Jovic, Brad Ferrier, Matt Pyle,
Geoff Manikin, Bo Yang, Jamie Wolff (DTC) and Brian Etherton (DTC)
http://www.emc.ncep.noaa.gov/mmb/SREF/SREF.html
Acknowledgements:
Ensemble Team: Yuejian Zhu, Yan Luo and Bo Cui
Mesoscale Branch: Julia Zhu, Eric Rogers, Perry Shafran and Ying Lin
HPC:
Dave Novak and Faye Barthold
AWC:
David Bright and Amy Harless
IBM:
Jim Abeles
NCO:
Xiaoxue Wang, Chris Magee, Becky Cosgrove , Carissa Klemmer
Motivation of this talk: not simply to introduce
the new SREF or show how great it is but to
facilitate R2O and O2R by providing you the
information “what we have and what we need”
Research  Operation

Research
Aug. 21, 2012 SREF upgrade
•
Model Change
1. Model adjustment (eliminate Eta and RSM legacy models and add new NEMS-based NMMB
model)
2. Model upgrade (two existing WRF cores from v2.2 to version 3.3)
3. Resolution increase (from 32km/35km to 16km)
4. All models run with 35 levels in the vertical and 50 mb model top.
•
IC diversity improvement
1. More control ICs (NDAS -> NMMB, GDAS -> NMM, RAP blended @ edges w/GFS -> ARW)
2. More IC perturbation diversity (blend of regional breeding and downscaled ETR)
3. Diversity in land surface initial states (NDAS, GFS, and RAP).
•
Physics diversity improvement
1. More diversity of physics schemes (flavors from NAM, GFS, NCAR and RAP)
•
New capabilities of post-processing & product generation
1. precipitation bias correction (individual members and ensemble mean)
2. clustering and associated mean/prob/spread within a cluster
3. member performance ranking (different weights for different members)
4. downscaling to 5km using RTMA and associated ensemble products.
• New ensemble products
1. max/min, mode, 10-25-50-75-90% forecasts
2. probs of severe thunderstorm, lightning, dry lightning, fire weather (SPC) as well as LLWS,
composite reflectivity, echo top, ceiling and visibility
3. addition of hourly ensemble product output from 1-39 hr.
4. ensemble mean bufr
5. a new 16km output grid covering North America (g132)
3
Evaluation of SLP (old SREF vs. new SREF, Oct. 23 – Dec. 31, 2011)
5
14
Ens mean fcst: RMSE
12
4
10
3
old
6
new
4
1
3
15 27 39 51 63 75 87
-0.1
-0.2
-0.3
120
Prob fcst: RPSS (12km NAM as ref)
1 3 5 7 9 11 13 15 17 19 21
Prob fcst: reliability diagram
100
old
new
no skill
15 27 39 51 63 75 87
80
60
40
old
new
perfect
20
0
0
9.5
19
28.6
38.1
47.6
57.1
66.7
76.2
85.7
95.2
0
new
0
0.2
0.1
old
2
0
0.3
Outlier of old SREF = 21.6% (to miss truth)
Outlier of new SREF = 15.7% (to miss truth)
8
2
0.4
Ens spread: Rank Histogram
4
Ensemble mean forecast and ensemble spread (24h-apcp,
against CCPA, Oct. 23 – Dec. 31, 2011)
0.35
20
18
0.3
16
0.25
14
0.2
12
0.15
old
0.1
Too much light precipitation at dryend but less chance to miss heavier
precipitation events at heavier precip
end
new
10
old
8
new
6
4
0.05
2
0
0
0.01" 0.1" 0.25" 0.5" 0.75" 1.0"
1
3
5
7
9 11 13 15 17 19 21
5
Strong cold bias in 2mT in warm season (old SREF vs. new SREF)
(b) Cold season (Oct. 23 – Dec. 31, 2011)
(a) Warm season (Jun. 15-Jul. 15, 2012)
Strong cold bias
(due to wet GFS?)
18
16
14
Outlier of old SREF = 18.81% (to miss truth)
Outlier of new S REF = 19.7% (to miss truth)
12
10
8
12
10
Outlier of old SREF = 17.3% (to miss truth)
Outlier of new SREF = 14.3% (to miss truth)
8
old
new
old
6
new
4
6
4
2
2
0
0
1 3 5 7 9 11 13 15 17 19 21
1
3
5
7
Ens spread: Rank Histogram
9 11 13 15 17 19 21
6
Dec. 1, 2011 West Coast HighWind Event (R. Grumm)
Old SREF
obs
New SREF
7
Day 1 forecast of 24h snow amount ending at 01/13/2012
Obs
Old SREF Mean
4km AFWA ENS Mean
HPC
New SREF Mean
8
18h-forecast of prob of CAPE > 4000 J/kg (valid at 03z, 6/30/2012)
Old SREF (underestimated)
New SREF (better)
9
PQPF of Old SREF (upper, underestimated)
vs. New SREF (lower, much improved)
Slow-moving Hurricane Debby-induced
heavy rain over Florida: observed 30hr
Accumulation (Rich Grumm)
>150mm
>200mm
Old SREF
obs
>150mm
>200mm
New SREF
10
Hurricane Issac track and rain forecasts
(made landfall in the west of New Orleans around 20z Aug. 29, 2012
11
What we have and what we need under the current
framework of NCEP SREF in the following aspects?
*IC perturbation including LBC and land surface initial states
*Model/physics perturbation
*Post processing/calibration
*Ensemble products and product-generator
*Other regional ensemble systems
*Verification (not to be touched here)
http://www.emc.ncep.noaa.gov/mmb/SREF/SREF.html
12
Initial condition/LBC/land surface initial states
What we have:
(1) Multi-analysis: GFS, NAM and RR analyses
(2) Mixed IC perturbations:
*regional bred vector (7 nmmb members)
*global ETR (7 wrf_arw members)
*blended perturbation of “smaller-scale bred vector + largerscale ETR” (7 wrf_nmm members)
(3) Various LBCs from global ensembles
(4) Various land surface initial states from NAM, GFS, RR
analyses
What we need:
(1) Refining IC perturbations such as 3D masking
(2) Connecting to NDAS system by using EnKF perturbations
(2) LBC perturbation scheme by better coupling with global model
(3) Direct perturbing land surface initial states
13
Model and physics
What we have:
(1) Multi-model: NMMB, WRF_NMM, WRF_ARW
(2) Multi-physics: various flavors from NMM, NCAR, GFS, RR (DTC helped to
test some schemes)
*see details in the table next slide
(3) Stochastic parameterization (Teixeira and Reynolds 2008) in
NMMB model is in place but not turned on
(4) Stochastic kinetic energy backscatter (SKEB) scheme in WRFARW model is in place but not turned on
What we need:
(1) Evaluating “the stochastic parameterization scheme” in NMMB
(2) Adding the same scheme to NMM and ARW models
(3) Needing to significantly speed up the SKEB in ARW model (Judith Berner)
(4) Adding SKEB to NMMB and NMM models (DTC)
(5) Testing other stochastic schemes
(6) Real question is to see if any stochastic physics scheme can really outperform
multi-model and multi-physics approaches?
14
Member
(Model)
IC
physics
List of the
physics schemes
IC
perturb.
Land surface
conv
mp
lw
sw
pbl
Sfc layer
stochastic
model
BMJ
FER
GFDL
GFDL
MYJ
MYJ
no
NOAH
SAS
GFS
GFDL
GFDL
GFS
MYJ
no
NOAH
BMJ
WSM6
GFDL
GFDL
MYJ
MYJ
NOAH
BMJ
FER
(new
Eta)
GFDL
GFDL
MYJ
M_Obuhov (Janjic Eta) no
NOAH
SAS
FER
(new Eta)
GFDL
GFDL
MYJ
M_Ouhov
(janjic Eta)
no
NOAH
KF
(new
Eta)
FER
(new
Eta)
GFDL
GFDL
MYJ
M_obuhov
(janjic Eta)
no
NOAH
KF
(new
Eta)
FER
(new
Eta)
GFDL
GFDL
MYJ
M_obuhov
(Janjic Eta)
no
NOAH
arw_n1
arw_p1
arw_n2
BMJ
FER
(new
Eta)
GFDL
GFDL
MYJ
M_obuhov
(Janjic Eta)
no
NOAH
arw_p2
arw_n3
BMJ
FER
(new eta)
GFDL
GFDL
MYJ
M_Obuhov
(Janjic Eta)
no
NOAH
nmmb_ctl
nmmb_n1
nmmb_p1
nmmb_n2
nmmb_p2
nmmb_n3
nmmb_p3
nmm_ctl
NDAS
GFS
BV
Blend
nmm_n1
nmm_p1
nmm_n2
nmm_p2
nmm_n3
nmm_p3
arw_ctl
arw_p3
RAP
ETR
initial
perturb.
NAM
no
GFS
no
RAP
no
15
Post processing and calibration
What we have:
(1) Decaying-average method for bias correction for basic
atmospheric variables (1st moment only)
(2) Frequency-matching method for precipitation bias correction (1st
moment only)
(3) Downscaling of surface variables from 16km to 5km by
applying the difference between lower-res and higher-res
analysis (DTC helped to test the code)
What we need:
(1) 2nd moment (spread) calibration (Bruce Veenhuis/MDL,
Decaying-average Bayesian Model Averaging)
(2) Bias correction of model variables directly on model native grid
(history file), so everything else produced by model post
thereafter will be automatically bias corrected
(3) Higher-moment: e.g. calibrating probability as well as estimating
16
uncertainty in probability (probability of probability)?
Performance of the downscaled 5km SREF (verified against
RTMA, 6/18/12 – 7/16/12)
Raw
Bias corrected Downscaled
1.4
2.5
1.2
2
1.5
1
SREF-RMSE
RAW-RMSE
BC-RMSE
SREF-Spread
RAW-Spread
BC-Spread
0.5
0
12
24
36
48
60
72
Spread/RMSE ratio
RMSE and Spread
1
0.8
0.6
0.4
SREF-Spread/RMSE
0.2
RAW-Spread/RMSE
BC-Spread/RMSE
0
12
24
84
Forecast hours
36
48
60
72
Forecast hours
T2m (ens mean)
17
84
Bias correction can effectively remove over-predicted
light precipitation and enhance under-predicted
heavier precipitation
16km SREF mean (raw)
16km SREF mean (bias corrected)
24
18
Ensemble products and product-generator
What we have:
(1) mean, spread and probability
(2) max/min, mode, 10-25-50-75-90% forecasts
(3) clustering
(4) member performance ranking (different weights for
different members)
What we need?
(1) Further testing of weighted ensemble mean
(2) Probability-matching mean (DTC visitor program)
(3) Neighborhood probability (DTC visitor program?)
(4) Extreme weather index probability
(5) Special products for wind energy and dispersion
uncertainty modeling
19
New ensemble products for aviation weather
Mean ceiling
Mean visibility
Low ceiling prob
Low visibility prob
New ensemble products for convection and fire weather (SPC)
Severe thunder
dry lightning
lightning
fire weather threat
Ensemble Clusters
22
Individual member’ performance ranking (weights for
each members): Du and Zhou 2011 MWR
13
Ensemble mean
Verified best
Verified worst
Individual mean
Predicted best
Predicted worst
RMSE
11
9
7
5
12/2009
01/2010
02/2010
03/2010
04/2010
05/2010
06/2010
07/2010
08/2010
09/2010
23
Other regional ensemble systems at NCEP
What we have:
(1) Time-lagged North America Rapid Refresh Ensemble (NARRETL) for aviation: 12km, 10 members, hourly update, 12hr length
(2) Dynamically downscaled Hi-Res Ensemble Forecast (HREF)
system: 5km, 44 members, 12hrly cycle, 48hr length (Dualresolution hybrid ensembling method, Du 2004)
(3) Time-lagged 3km-HRRR-based Storm-Scale Ensemble Forecast
(SSEF–TL) for convection, aviation and dispersion (under
construction to replace and combine with HREF)
What we need:
(1) Learning from OU/NSSL/SPC SSEF and AFWA 4km-EPS
experiences
(2) Bigger and faster super-computer to actually run a 3km Hi-Res
Rapid Refresh based SSEF system to replace HREF and
HRRRE-TL as well as NARRE-TL
(3) NAEFS_LAM by combining SREF with Canadian REPS
24
We have moved from the old to the new building!
(the move-in date: 8/20/12; the SREF implementation date: 8/21/12)
Next generation SREF
New SREF
Convection-explicit, cloudresolving, rapid-update-cycle
storm-scale ensemble
prediction system directly
coupled with data assimilation
Old SREF
25