Aims and Requirements for Ensemble Forecasting By T.N.Palmer ECMWF
Mar 27, 2015
Aims and Requirements for Ensemble Forecasting
By
T.N.Palmer
ECMWF
A Brief History of Ensemble Prediction for Weather/Climate
Monthly forecasting
Medium Range Seasonal-to-Interannual
Short range Climate Change
1980s
1990s
2000s
“Roots of ensemble forecasting” J.M Lewis. Mon Wea Rev, 133, 1865 (2005)
First operational probabilistic ensemble forecast?
Used in Met Office commercial operations.
Scientific Basis for Ensemble Prediction
In a nonlinear dynamical system, the finite-time growth of initial uncertainties is flow dependent.
Lorenz (1963): prototype model of
chaos
October 1987!
is a nonlinear system
Since is a nonlinear function of
( )
dXF X
dtd X dF
X J Xdt dXF X
J J X
In a nonlinear system, finite-time predictability is a function of initial state
EPS spread/Error
12UTCEurope Lat 35.0 to 75.0 Lon -12.5 to 42.5
500hPa GeopotentialTime series curves
MAR2005
APR MAY JUN0
20
40
60
80
100
120
ENSTD T+24 Mean forecast
ENSTD T+48 Mean forecast
ENSTD T+144 Mean forecast
ENSMN T+24 Root mean square error forecast
ENSMN T+48 Root mean square error forecast
ENSMN T+144 Root mean square error forecast
ECMWF Ensemble Forecasts of Katrina
26 Aug 0z26 Aug 12z
27 Aug 0z 27 Aug 12z
Who is coming? Pthreshold
Queen 1%
Mayor
Mother-in-law
Mates from the pub
10%
30%
70%
Decision: Rent marquee if P>20%
Queen Mayor Mother
-in-lawMates from pub
From ECMWF web site.
Weather Roulette• London-Heathrow, 2m temperature
• 2002: training data for dressing
• 2003: test data
• odds: set by dressed T511 forecast
• bets: placed by best member dressed EPS
• start capital: £1 (re-invest all money, unlimited stakes)
odds(bin) = 1 / prob_hr(bin)bets(bin) = prob_eps(bin) * capital(t-1)
Daily winnings:
win(t) = odds(bin_v) * bets(bin_v) – capital(t-1)
= (prob_eps(bin_v)/prob_hr(bin_v) – 1) * capital(t-1)
Collaboration with L.Smith, LSE
Weather Roulette
Days in 20030 50 100 150 200 250 300 3500
20
40
Win
nin
gs
[lo
g_1
0 £]
168h lead time
Collaboration with L.Smith, LSE
Weather Roulette
Lead time [days] 1 2 3 4 5 6 7 8 9 100
25
50
Win
nin
gs
[lo
g_1
0 £]
Bootstrapping Results
Collaboration with L.Smith, LSE
DEMETER-based PDFs of malaria incidence for Botswana (forecasts made 5 months in advance
of epidemic; Thomson et al 2005)
5 years with lowest observed malaria incidence
5 years with highest observed malaria incidence
Why No Ensembles on the TV Weather
Forecasts?
956
AN 19871016, 06GMT
979
EPS Cont FC +66 h
984
- mem no. 1 of 51 +66 h
963
- mem no. 2 of 51 +66 h
968
- mem no. 3 of 51 +66 h
978
981
- mem no. 4 of 51 +66 h
979
- mem no. 5 of 51 +66 h
962
- mem no. 6 of 51 +66 h
988
- mem no. 7 of 51 +66 h
966
- mem no. 8 of 51 +66 h
969
- mem no. 9 of 51 +66 h
981
984
- mem no. 10 of 51 +66 h
964
- mem no. 11 of 51 +66 h - mem no. 12 of 51 +66 h - mem no. 13 of 51 +66 h
965
979
- mem no. 14 of 51 +66 h
990
- mem no. 15 of 51 +66 h
965
- mem no. 16 of 51 +66 h
976
- mem no. 17 of 51 +66 h
970
- mem no. 18 of 51 +66 h
984 - mem no. 19 of 51 +66 h
962
- mem no. 20 of 51 +66 h
961
- mem no. 21 of 51 +66 h - mem no. 22 of 51 +66 h
966
- mem no. 23 of 51 +66 h
970
979
- mem no. 24 of 51 +66 h
975
982
- mem no. 25 of 51 +66 h - mem no. 26 of 51 +66 h
964
- mem no. 27 of 51 +66 h
979 - mem no. 28 of 51 +66 h - mem no. 29 of 51 +66 h
967
- mem no. 30 of 51 +66 h
964
980
- mem no. 31 of 51 +66 h
983
- mem no. 32 of 51 +66 h
980
- mem no. 33 of 51 +66 h
974
- mem no. 34 of 51 +66 h
972
- mem no. 35 of 51 +66 h
981
- mem no. 36 of 51 +66 h
964
972 - mem no. 37 of 51 +66 h
988
- mem no. 38 of 51 +66 h
978
978
- mem no. 39 of 51 +66 h
960
- mem no. 40 of 51 +66 h
985
988
- mem no. 41 of 51 +66 h
977
980
- mem no. 42 of 51 +66 h
979
986
- mem no. 43 of 51 +66 h
976
- mem no. 44 of 51 +66 h
980
- mem no. 45 of 51 +66 h
958
- mem no. 46 of 51 +66 h
968
- mem no. 47 of 51 +66 h
987
- mem no. 48 of 51 +66 h
963
- mem no. 49 of 51 +66 h
989
- mem no. 50 of 51 +66 h
MLSP 66-hour forecasts, VT: 16-Oct-1987, 6 UTC
TL399 EPS with TL95, moist SVs
Probability of Beaufort force 12 winds 6-12am October 16th 1987
“Weather forecasts are inevitably uncertain, sometimes more so than others. We now run our forecast models many times with slightly different starting conditions to assess the uncertainty in the forecasts. Press the red button on your remote control to see an estimate of the expected accuracy of the forecast for some of the main cities in the UK.”
Aims of Ensemble Forecasting
•To enhance (substantially) the value of numerical weather and climate forecasts by quantifying the flow-dependent uncertainty in the forecast
•To enhance the credibility of weather and climate forecasts, thereby allowing our profession to gain the respect of the public
Requirements for Ensemble Forecasting
•A better theoretical understanding of the role of error made in truncating/parametrizing the underlying PDEs of climate, on both initial uncertainty and forecast model uncertainty
•Much much much bigger computers (resolution, ensemble size and model complexity are all important)
•A recognition amongst media forecasters that uncertainty is an intrinsic part of the science of weather and climate prediction..and that the public will respect us more if we are more open about uncertainty
•A recognition amongst BBC TV producers that use of ensemble forecast information on weather forecasts can inform, educate and entertain the viewing public and is something worth giving more effort to.
“He believed in the primacy of doubt; not as a blemish upon our ability to know, but as the essence of knowing”Gleick (1992) on Richard Feynman’s philosophy of science.