Transcript EnsMOS_pres

Using
Ensemble Model Output Statistics
to Improve 12-Hour
Probability of Precipitation Forecasts
John P. Gagan
NWS Springfield, MO
Chad Entremont
NWS Jackson, MS
NWS Versus MEX
12-Hour PoP Forecasts (Separated Wet from Dry)
0.7
Wet PoP
0.6
TROUBLE
Probability
0.5
0.4
0.3
0.2
0.1
Dry PoP
GOOD
0
1
2
3
4
5
6
7
8
9
10
11
12
Forecast Period
WFO Wet PoP
MEX Wet PoP
WFO Dry PoP
MEX Dry PoP
13
14
The Problem
• Dry Bias
– Improvement noted with No Precip
– Forecaster not as “wet” as GFS MOS when
there is Precip
• National problem
– Almost all areas exhibit same tendencies
– Issues in both cold and warm seasons
The Problem (con’t)
• PoP definition
– The probability of occurrence of
measurable precipitation (0.01 inch) at a
given point for each 12-hour period
through Day 7.
• The Interpretation
– In spite of a straightforward definition, it’s
as unique as the individual asked
• “PoPs ‘look’ too high today”
• “It’s not going to rain that much, so I’m
lowering PoPs”
• “I never go likely beyond 48 hours”
A Generic Forecast ‘In Words’
• “Models increasing PWs to 200% of
normal”
• “High Θe air being pumped into region by
50kt LLJ”
• “Area in right entrance region of ULJ”
• “Large area of rain and embedded
thunderstorms will move over the area
today”
• And so on…
A Generic Forecast ‘By Numbers’
• MOS PoP Forecasts
– MAV – 90%
– MET – 85%
– Ensemble MOS – 80%
• Forecaster’s PoP
– 70% area wide
What Happened?
• Numerous reasons why it WILL rain
• Yet, the forecast is drier than MOS
• Why?
– Mistrust/misunderstanding of MOS?
– Lack of understanding of the 12-hr
PoP?
– Reasons vary by individual
– The main issue - CONFIDENCE
A Solution
• Ensemble MOS
– Started April 2001
– Currently a 16-member suite
• Operational MEX
• Control Run
• 14 Perturbations
– Run 1-time per day (00z issuance)
• A bulletin is created showing the
Max/Min/Avg of MOS output
A Solution (con’t)
• Use the Ensemble Average PoP as a
means to improve PoP forecasts
– DO NOT use the ensemble average value
as an EXPLICIT forecast
– Use the ensemble average value as
CONFIDENCE factor
• The higher the ensemble
average, the more confidence
in precip occurrence
Data Manipulation
• This investigation is for the COLD
SEASON ONLY!
– October to April
• Data collected from Oct 2003 – Apr
2006
• Investigated 6 sites
– SGF CWFA – KSGF, KVIH
– JAN CWFA – KGWO, KTVR, KJAN,
KMEI
Data Manipulation (con’t)
• ~ 4000 data points collected
– Stratified by rain/no rain
• Periods 1-10 studied (Days 1-5)
• Graphs produced to highlight rainfall
frequency for a given value of the
ensemble average PoP
• Ensemble Average PoP is NOT used
as a PoP but a confidence factor
Cases
Rain / All Cases
Periods 1 and 2
371 / 390
891 / 1046
1
1127 / 1501
95.1%
85.2%
0.9
75.1%
0.8
Percent of Time
it Rains
0.7
0.6
0.5
0.4
0.3
0.2
0.1
7 / 3172
0.2%
43 / 4496
1.0%
0
<5
<15
>=40
>=55
Ensemble Average PoP
>=80
Cases
Rain / All Cases
Periods 3 and 4
24 / 25
613 / 715
96.0%
1
85.7%
966 / 1297
0.9
75.5%
0.8
Percent of Time
it Rains
0.7
0.6
0.5
0.4
0.3
0.2
5 / 2523
66 / 4372
0.2%
1.5%
0.1
0
<5
<15
>=40
>=55
Ensemble Average PoP
>=80
Cases
Rain / All Cases
Periods 5 and 6
397 / 471
1
776 / 1084
84.3%
0.9
71.6%
0.8
Percent of Time
it Rains
0.7
0.6
0.5
0.4
0.3
7 / 1950
0.2
0.4%
99 / 4149
2.4%
0.1
0
<5
<15
>=40
Ensemble Average PoP
>=55
Cases
Rain / All Cases
Periods 7 and 8
211 / 263
1
891 / 1046
80.2%
0.9
68.5%
0.8
Percent of Time
it Rains
0.7
0.6
0.5
0.4
0.3
19 / 1217
109 / 3770
0.2
1.6%
2.9%
0.1
0
<5
<15
>=40
Ensemble Average PoP
>=55
Cases
Rain / All Cases
Periods 9 and 10
1
0.9
221 / 344
0.8
559 / 992
Percent of Time
it Rains
0.7
64.2%
56.4%
0.6
0.5
0.4
0.3
55 / 1823
143 / 3253
0.2
3.0%
4.4%
0.1
0
<10
<15
>=35
Ensemble Average PoP
>=45
Using Ensemble PoP in Real Time
• Using the ensemble average alone does well
• However, using this in tandem with the full
suite of models/guidance is best
– SREF – (Probabilities from SPC web page)
– Other global models
– Mesoscale models
• The more datasets that say “YES” should
increase confidence and result in a better
quality PoP
Observations and Further Study
• Watch the day-to-day trend of the
ensemble average PoP
– If the value increases for a particular period,
confidence increases
– Should be able to hone in on 1-3 12-hour
periods (“windows of opportunity”)
• Watch for MEX PoP values LESS THAN
the ensemble average
– Observation has shown that it does not rain
as often
Does It Work?
A quick look at verification
from KJAN
KJAN Versus MEX
PoP Forecasts (Separated Wet from Dry)
Jan 2004 to Apr 2005
0.9
0.8
Probability
0.7
Dry Bias Prominent
0.6
0.5
0.4
0.3
0.2
0.1
0
1
2
3
4
5
6
7
8
9
10
11
12
Forecast Period
JAN Wet PoP
MEX Wet PoP
JAN Dry PoP
MEX Dry PoP
13
14
KJAN Versus MEX
PoP Forecasts (Separated Wet from Dry)
Oct 2005 to Apr 2006
0.9
0.8
Probability
0.7
Dry Bias Eliminated
0.6
0.5
0.4
0.3
0.2
0.1
0
1
2
3
4
5
6
7
8
9
10
11
12
Forecast Period
JAN Wet PoP
MEX Wet PoP
JAN Dry PoP
MEX Dry PoP
13
14
Questions, Comments?
• If you are interested in this study, we’d like
to hear your opinions
– [email protected]
– [email protected]