page 1© crown copyright 2004 presentation to ecmwf forecast product user meeting 16th june 2005
TRANSCRIPT
© Crown copyright 2004 Page 1
Presentation to ECMWF Forecast Product User Meeting
16th June 2005
© Crown copyright 2004 Page 2
Met Office use and verification of the monthly forecasting system
Bernd Becker, Richard Graham.
Met Office monthly forecast suite
Products from the Monthly Outlook
Standard Verification System
© Crown copyright 2004 Page 3
Monthly Forecasting System
Coupled ocean-atmosphere integrations: a 51-member ensemble is integrated for 32 days every week.
Atmospheric component: IFS with the latest operational cycle 29r1 and with a TL159L40 resolution (320 * 161)
Oceanic component: HOPE (from Max Plank Institute) with a zonal resolution of 1.4 degrees and 29 vertical levels
Coupling: OASIS (CERFACS). Coupling every ocean time step (1 hour)
Perturbations:
Atmosphere: Singular vectors + stochastic physics
Ocean: SST perturbations in the initial conditions + wind stress perturbations during data assimilation.
Hindcast statistics:
5-member ensemble integrated over 32 days during the past 12 years.
Representing a 60-member ensemble.
Running every week
© Crown copyright 2004 Page 4
HC min HC max
qb1 qb 2 qb 3
qb 4
FXmeanFXmin FXmax
qm1 qm2 qm4qm3 qm5
FORMOST global PDF data
ForecastHind cast
Properties of quintile PDF
well below below normal above well above
© Crown copyright 2004 Page 5
FORMOST Monthly Outlook
© Crown copyright 2004 Page 6
Example UK 12-18 day temperature forecast for 10 climate districts
Deterministic forecast
(based on most probable category or ensemble mean)
Probability forecast
Verification
“….a sudden change to below or well-below average temperatures is expected…”
(Forecast text issued 11th Feb for week 21-27 of Feb)
© Crown copyright 2004 Page 7
Example global capability tercile probability forecast – Europe, days12-18
Verification (ECMWF operations)
P(abv)
P(avg)
P(blw)
PrecipTmean
© Crown copyright 2004 Page 8
WK 3&4
WK 1 Obs. T anom. 24-30. Jan
WK 2
© Crown copyright 2004 Page 9
Verification May 2002 - April 2005
93 forecast/ observation pairs of Temperature and Precipitation
Verifying Observations: ECMWF short range (12-36 hrs) forecasts over the
period
Global Forecasts: Relative Operating Characteristics for quintile
forecast Reliability Diagram Brier skill score decomposition
© Crown copyright 2004 Page 10
Relative Operating Characteristics
ROC measures the ability of the forecast to discriminate between two
alternative outcomes.
measuring resolution. It is not sensitive to bias in the forecast, so
says nothing about reliability. A biased forecast may still have good resolution and produce a good ROC curve, which
means that it may be possible to improve the forecast through
calibration. The ROC can thus be considered as a measure of potential
usefulness.
The ROC is conditioned on the observations (i.e., given that bin X
occurred, what was the corresponding forecast?)
The area under the ROC curve indicates skill above climatology, when
> 0.5
POFD=
FA/(FA+CR)
POD=
H/(H+M)
3
cold
warm
© Crown copyright 2004 Page 11
ROC Score for Temperature
well below normal 19 to 32
days ahead
ROC Map
© Crown copyright 2004 Page 12
Brier Score = RMS ensemble forecast 1/N Sum_i (P_k – O_k)^2
BS = Reliability – Resolution + Uncertainty
Reliability: conditional Bias, weighted vertical distance of the reliability curve to the diagonal. The flatter the less resolution.
Resolution: difference between fx pdf and expectation, variance (HR). The larger (better), the more often the fx pdf is steeper and narrower than the climatological pdf. (Flat sharpness plot)
Uncertainty: difference between obs pdf and sample average pdf, variance(H/total), .2(1-.2)=0.16 always > Resolution
Bracketed Terms:
BSS= 1 – BS/BS_clim
BS_clim= REL(=0) - Res(=0) + Unc
Brier ScoreHR=H/(H+FA)
3
© Crown copyright 2004 Page 13
Monthly Verification (Europe): TmeanROC & reliability, all seasons, days 12-18
• Based on 93 operational forecasts (all seasons)
• Verification data = ECMWF T+12-36
POD=
H/(H+M)
HR=H/(H+FA)2 2
© Crown copyright 2004 Page 14
Monthly Verification (Europe): PrecipROC & reliability, all seasons, days 12-18
POD=
H/(H+M)
POFD=
FA/(FA+CR)
HR=H/(H+FA)2 2
© Crown copyright 2004 Page 15
Conclusion
The monthly forecasts model runs are produced at ECMWF, products are derived at the Met Office, operationally.
Standardised Verification system (SVS) for Long-range Forecasts (LRF) is beginning to take shape.
Forecasts for day 19-32 are as useful as climatology.
Predictions of Quintile 1 and 5 are more skilful than of Quintiles 2 to 4
Europe is a difficult region to predict at long time range. Weather uncertainty imposes a fundamental limit on the sharpness of the probabilistic forecast.
The Monthly Outlook is a powerful tool to provide forecast guidance up to a month ahead in many areas.
© Crown copyright 2004 Page 16
THANK YOU!
Richard Graham
Margaret Gordon
Andrew Colman
© Crown copyright 2004 Page 17
The Met Office
We are one of the world's leading providers of environmental and weather-related services.
Our solutions and services meet the needs of many communities of interest…from the general public, government and schools, to civilaviation and almost every industry sector around the world.
© Crown copyright 2004 Page 18
Desirable Attributes of Scores
Equitable:
Equitable without dependence on the forecast distribution:
Rewards for correct forecasts inversely proportional to their event frequencies
Penalties for incorrect forecasts directly proportional to their event frequencies
Random forecasts and constant forecasts have the same (0) skill.
Note: 2 & 3 imply that all information in the contingency table is taken into account.
© Crown copyright 2004 Page 19
Recap: Post processing and Products
Data Volume reduction before transfer to The Met Office: Calculate
1. Tercile/Quintile boundaries from the Hindcast ensemble2. Tercile/Quintile populations from the Forecast ensemble3. Maximum, Mean and Minimum from Forecast and from Hindcast4. Forecast Tercile/Quintile averages5. Average in time to week 1, 2 and 3&4.
UK Forecast: http://www.bbc.co.uk/weather/ukweather/monthly_outlook.shtml1. Interpolation to points representing UK climate regions2. Calibration with historical UK climate region observations3. Interpretation of the Histogram, Ensemble mean or Mode in cases with large spread,
• derive deterministic forecast tercile/quintile4. Mapping Tercile/Quintile average onto calibration PDF
• to derive deterministic forecast value
Global Forecast:1. Tercile/Quintile probabilities2. Calibrate by overlaying Tercile/Quintile boundaries derived from 1989 – 1998 ERA40
data
© Crown copyright 2004 Page 20
Key Points
Monthly Outlooks are a split process between ECMWF and Met Office
The Monthly Outlook is operational
Standardised Verification system is shaping up
© Crown copyright 2004 Page 21
Most probable Quint
Low spread, Ensemble mean a good “best estimate”, high confidence.
High Spread, ensemble mean wrong in 80% of all cases.
High Spread, delta probability > 5%.Most probable quint “best estimate”.Hedge by use of above/below quint.
© Crown copyright 2004 Page 22
FORMOST Verification Week 2
Quint boundaries
ObservedDeterminist
ic
t
© Crown copyright 2004 Page 23
FORMOST Verification Week 3&4, UK average
t
© Crown copyright 2004 Page 24
FORMOST Verification Week 3&4, UK average
t
© Crown copyright 2004 Page 25
Monthly forecast: 3 category probabilities
prob. of above
prob. of average
prob. of below
Issued: 9th June 2005
valid: 20th to 26th June
temperature precipitation
© Crown copyright 2004 Page 26
Grid point diagnostics
Properties of a contingency tableper grid point
wan 5
QUINT 4
n 3
bn 2
wbn 1
H M FA CR
100 % probability
90
80
70
60
50
40
30
20
10
0
0
1
1
3
5
8
13
20
23
1
1
2
3
7
8
11
13
20
29
Hit : Q =Qobs & P(Q)>= Pthresh
Miss: Q =Qobs & P(Q) < Pthresh
False Alarm:
Q != Qobs & P(Q)>= Pthresh
Correct rejection: Q !=Qobs &
P(Q)<Pthresh
Stratify by magnitude of the
probability at each grid point
POD = H / (H+M) conditioned on Observations
POFD=FA / (FA+CR)
Hit Rate = H / (H+FA) conditioned on Forecasts
© Crown copyright 2004 Page 27
EUROPE T
HR=H/(H+FA)
19 lead time
2 weeks period
3
3
POD=
H/(H+M)
© Crown copyright 2004 Page 28
EUROPE PHR=H/(H+FA)
19d lead time
2 weeks period
3
3
POD=
H/(H+M)