1 Mt. Mansfield, VT
Dec 16, 2015
1Mt. Mansfield, VT
2
Aka….New England Days 4-7 Forecast Test
Paul A. SissonNOAA/NWS, Weather Forecast Office, Burlington,
Vermont (BTV)
Joseph DellicarpiniNOAA/NWS, Weather Forecast Office, Taunton,
Massachusetts (BOX)
Michael EksterNOAA/NWS, Weather Forecast Office, Gray, Maine
(GYX)
Todd FoisyNOAA/NWS, Weather Forecast Office, Caribou, Maine
(CAR)
David Radell and Jeff WaldstreicherNOAA/NWS, Eastern Region Headquarters, Bohemia,
New York (ERH)
Mt. Mansfield, VT
3
The real workers
• Matt Belk BOX• Conor Lahiff BTV• Margaret Curtis GYX• Roman Berdes CAR
Acknowledgements• The NWS Central Region Blender
team• Tim Barker, SOO, BOI - BOIVerify
4
Outline
• Background• Motivation• The Experiment• Verification Results• Summary
“The Blend is your friend”
5
Background• Many Studies on forecast/model consensus in general
agreement that the consensus forecast tends to be the best forecast– Gyakum (1986), Fritsch et al (2000), Roebber (2014) etc.
Baars and Mass (2005)
Consensus/Weighted MOS – “competitive or superior to human forecasts at nearly all locations”
“Human forecasts are most skillful compared to MOS during the first forecast day and for periods when temperatures differ greatly from climatology.”
6
Motivation
• Work efficiently• Improve Accuracy• Improve Consistency• Focus on what we do best – (sig departures from climo and short
term)
Note: Don’t reinvent the wheel“The Blend is your friend”
7
Consistency Problems
8
The Days 4-7 Forecast Experiment
• Oct 2013 - Mar 2014• All offices start with same model
initializations and blends• Forecaster Surveys and Verification• Does blend outperform Gridded MOS?• Is the forecast more consistent?• Does the method allow forecasters to
be more efficient and allow time to do important things?
9
Nomenclature
• CONS = Consensus • A consensus data set is
calculated by combining/averaging a “list” of guidance data sources.
10
BC = Bias Correction
• Bias corrections are run on individual guidance sources using BOIVerify software.
• BCCONS data set is generated by bias correcting the individual components before forming the consensus.
• Using previous 14 day period BC
11
BLENDS
• Blends are a combination of various datasets of various weights.
Name Description
CMCnh Canadian Meteorological Center Global model Raw Model output
GFS40 NWS Global Forecast System (GFS: 40km) Raw Model output
ECMWFEuropean Centre Medium Range Forecast model
Raw Model output
SREF NWS Short Range Ensemble Forecast (mean) Raw Model output
NAM12 NWS North American Model (NAM: 12km) Raw Model output
NAMDNG5
NWS NAM downscaled to 5km grid Raw Model output
ADJMEXGFS40km ADJusted with GFS extended MOS point forecasts
Raw Model background with MOS
ADJMENGFS40km ADJusted with GFS ensemble mean MOS pt forecasts
Raw Model background with MOS
ADJECEECMWF ADJusted with ECMWF MOS pt forecasts
Raw Model background with MOS
ADJECMECMWF ADJusted with ECMWF ensemble mean MOS pt forecasts
Raw Model background with MOS
MOSG25 GFS Gridded Model Output Statistics (2.5km) GFS MOS
*Note: all forecasts mapped/downscaled to 2.5km grid
Forecast Databases
Name Description Databases
CONSAllConsensus Raw Models and MOS
CMCnh, GFS40, ECMWF, SREF, NAM12, NAMDNG5, MOSG25, ADJMEX, ADJMEN, ADJECE, ADJECM (equal weights)
BCCONSAll
Consensus of Bias-corrected Raw Model and MOS databases
CMCnhBC, GFS40BC, ECMWFBC, SREFBC, NAM12BC, NAMDNG5BC, MOSG25BC, ADJMEXBC, ADJMENBC, ADJECEBC, ADJECMBC
HPCGuideHuman Forecast by Weather Prediction Center
Human adjusted Blend
OfficialPrevious Weather Forecast Office forecast
Human adjusted Blend
Forecast Databases
14
SuperBlend
• Previous Forecast + latest blends
• Official (25%), • HPCGuide (25%) • CONSALL (25%) • BCCONSALL (25%)
• 50/50 Man Machine mix
15
Verification Results
• MaxT, MinT, T, Td, Wind Speed, PoP
• Use Real-Time Mesoscale Analysis adjusted by Observation as verification and for bias-correction
RTMA MaxT
Obs to ADJ RTMA MaxT
18
RESULTS
19
Max_T Min_T T Td Wind Speed
-10
-5
0
5
10
15
20
25
30All WFOs Days 4-7 Mean % Improvement
Over GMOS
OfficialHPCGuideSuperblendCONSALLBCCONSALL
% Im
prov
emen
t ove
r GM
OS
BOX: Nov-Mar 2014.GYX: Jan-Mar 2014.CAR: Nov/Mar2014.BTV: Oct.-Mar 2014.
20
AD
JECM
Raw
Blen
dA
llBle
ndSu
perB
lend
Offi
cial
AD
JECE
CON
SAll
AD
JECM
BCA
DJE
CEBC
BCCO
NSA
llBC
CON
SRaw
ECM
WFB
CCO
NSR
awCO
NSM
OS
AD
JMEX
BCCO
NSM
OS
CMCn
hBC
AD
JMEN
MO
SG25
CMCn
hEC
MW
FA
DJM
EXBC
HPC
Gui
deM
OSG
uide
MO
SG25
BCA
DJM
ENBC
MO
SGui
deBC
GFS
40BC
NA
M12
BCU
KMET
BCN
amD
NG
5BC
GFS
40EK
DM
OSB
CD
GEX
EKD
MO
SH
PCG
uide
BCU
KMET
DG
EXBC
NA
M12
Nam
DN
G5
2.003.004.005.006.007.008.009.00
10.0011.00
4.11 4.25
4.27
4.29 4.46 5.
11
5.19
6.79
Max T Avg MAE Periods 6-14 BTV CWA vs Obs
MAE
deg
F
Day 5 MinT M
OS
G2
5M
OS
Gu
ide
CO
NS
MO
SA
DJ
ME
XC
ON
SA
llG
FS
40
All
Ble
nd
HP
CG
uid
eO
ffic
ial
Su
pe
rBle
nd
MO
SG
uid
eB
CM
OS
G2
5B
CB
CC
ON
SM
OS
GF
S4
0B
CP
rev
iou
sB
CC
ON
SA
llB
CC
ON
SR
aw
EC
MW
FB
CA
DJ
ME
XB
CD
GE
XB
CC
MC
nh
BC
Ra
wB
len
dB
CC
ON
SS
ho
rtC
ON
SS
ho
rtC
ON
SR
aw
DG
EX
CM
Cn
hE
CM
WF
-6
-5
-4
-3
-2
-1
0
1
2
3
4
Day 5 MinT Bias BTV CWA
Series1
Day 4 PoP Reliability
Under forecast
Over forecast
23
Survey Results
• Compare Before, During, & After
• Before: Gridded MOS method
• After: SuperBlend method
Impact of SuperBlend on Days 4-8 Forecast Process
MEAN – 4.69
Mid-Test Survey Pre-Test Survey (MOSGuide)
MEAN – 4.48 MEAN – 3.45
Forecaster ModificationsWhat Doesn’t Work Well
SuperBlend
MOSGuide
SuperBlend PerformanceRegime Changes / Anomalous Conditions
SuperBlend PerformanceRegime Changes / Anomalous Conditions
MEAN – 4.64
Mid-Test Survey Pre-Test Survey (MOSGuide)
MEAN – 4.31 MEAN – 3.16
Forecaster Overall Evaluation of Approach
29
Consistency Mar 2013Less Than 80 %
80 - 90 % 90 - 95 % 95 - 100 % 100 %
30
QPF
31
SnowAmt
32
Summary
• Blends provide a more accurate and consistent starting point
• Forecasters survey show confidence in the method
• Allows more time for operations vs grid preparation
33
Bill Belichick, Head CoachNew England Patriots
Dec 14, 2009
Caveat: "Stats are for losers," "The final score is for winners."
34
Or….stated another way
Caveat: Stats are helpful
Clearly Communicating Accurate and Timely Weather Information is the goal.
35
The End
Name Description Databases
CONSRawConsensus of Raw Models
CMCnh, GFS40, ECMWF, SREF, NAM12, NAMDNG5 (equal weights)
BCCONSRaw
Consensus of Bias-corrected Raw Model databases
CMCnhBC, GFS40BC, ECMWFBC, SREFBC, NAM12BC, NAMDNG5BC (equal weights)
CONSMOSConsensus of MOS databases
MOSG25, ADJMEX, ADJMEN, ADJECE, ADJECM, EKDMOS, ADJMAV, ADJLAV, ADJMET (equal weights)
BCCONSMOS
Consensus of Bias-corrected MOS databases
MOSG25BC, ADJMEXBC, ADJMENBC, ADJECEBC, ADJECMBC, EKDMOSBC, ADJMAVBC, ADJLAVBC, ADJMETBC (equal weights)
AllBlend50/50 Blend of Official and CONSAll
Official, CONSAll
BCAllBlend50/50 Blend of Official and BCCONSAll
Official, BCCONSAll
RawBlend50/50 Blend of Official and CONSRaw
Official, CONSRaw
BCAllBlend50/50 Blend of Official and BCCONSRaw
Official, BCCONSRaw
Forecast Databases
Name Description Databases
SuperBlend Previous Forecast + latest blends Official (25%), HPCGuide (25%), CONSALL (25%), and BCCONSALL (25%)
CONSAll Consensus Raw Models and MOS CMCnh, GFS40, ECMWF, SREF, NAM12, NAMDNG5, MOSG25, ADJMEX, ADJMEN, ADJECE, ADJECM,, HPCERP (equal weights)
BCCONSAll Consensus of Bias-corrected Raw Model and MOS databases
CMCnhBC, GFS40BC, ECMWFBC, SREFBC, NAM12BC, NAMDNG5BC, MOSG25BC, ADJMEXBC, ADJMENBC, ADJECEBC, ADJECMBC
CONSRaw Consensus of Raw Models CMCnh, GFS40, ECMWF, SREF, NAM12, NAMDNG5 (equal weights)
BCCONSRaw Consensus of Bias-corrected Raw Model databases
CMCnhBC, GFS40BC, ECMWFBC, SREFBC, NAM12BC, NAMDNG5BC (equal weights)
CONSShort CONSRaw databases, Local WRF's and short term MOS
CMCnh, GFS40, ECMWF, SREF, NAM12, NAMDNG5, RUC13BC, HIRESWarwBC, HIRESWnmmBC, BTV4, BTV12, BTV6, ADJMAV, ADJMET, ADJECE, ADJLAV (equal weights)
BCCONSShort Consensus of Bias-corrected CONSShort databases
CMCnhBC, GFS40BC, ECMWFBC, SREFBC, NAM12BC, NAMDNG5BC, RUC13BC, HIRESWarwBC, HIRESWnmmBC,BTV4BC, BTV12BC, BTV6BC, ADJMAVBC, ADJMETBC, ADJECEBC, ADJLAVBC (equal weights)
CONSMOS Consensus of MOS databases MOSG25, ADJMEX, ADJMEN, ADJECE, ADJECM, EKDMOS, ADJMAV, ADJLAV, ADJMET (equal weights)
BCCONSMOS Consensus of Bias-corrected MOS databases
MOSG25BC, ADJMEXBC, ADJMENBC, ADJECEBC, ADJECMBC, EKDMOSBC, ADJMAVBC, ADJLAVBC, ADJMETBC (equal weights)
AllBlend 50/50 Blend of Official and CONSAll Official, CONSAll
BCAllBlend 50/50 Blend of Official and BCCONSAll Official, BCCONSAll
RawBlend 50/50 Blend of Official and CONSRaw Official, CONSRaw
BCAllBlend 50/50 Blend of Official and BCCONSRaw Official, BCCONSRaw
38
Max_T Min_T T Td Wind Speed
-15
-10
-5
0
5
10
15
20
25
30
WFO BOX Days 4-7 % Improvement Over GMOS
OfficialHPCGuideSuperblendCONSALLBCCONSALL
% Im
prov
emen
t ove
r GM
OS
39
Max_T Min_T T Td Wind Speed
-15
-10
-5
0
5
10
15
20
25
30
WFO BTV Days 4-7 % Improvement Over GMOS
OfficialHPCGuideSuperblendCONSALLBCCONSALL
% Im
prov
emen
t ove
r GM
OS
All data: Oct.-Mar.
40
Max_T Min_T T Td Wind Speed
-15
-10
-5
0
5
10
15
20
25
30
WFO CAR Days 4-7 % Improvement Over GMOS
OfficialHPCGuideSuperblendCONSALLBCCONSALL
% Im
prov
emen
t ove
r GM
OS
All data: Mar. 2014Min/Max: Nov.-Jan.
41
Max_T Min_T T Td Wind Speed
-15
-5
5
15
25
35
45
55
WFO GYX Days 4-7 % Improvement Over GMOS
OfficialHPCGuideSuperblendCONSALLBCCONSALL
% Im
prov
emen
t ove
r GM
OS
All data: Jan-Mar 2014.