Transcript PowerPoint

Lothar (T+42 hours)
5-Day ECMWF Ensemble Prediction of Typhoon Rusa
Global NWP models cannot predict extremes of precipitation:
need for coupling to LAMs
Extreme rainfall as a function of spatial scale (observational study:
Olsson et al, 1999)
EPS cannot resolve circulation features in this range (cf lack of k-5/3 spectrum in model)
ECMWF EPS – current operational
configuration
1. 51 members. TL255L40. Once per day (12z).
25 Initial + Evolved dry singular vectors
T42L40. 48 hour optimisation. Energy metric.
2. Stochastic physics
D+2
D+4
D+7
spread
control error
ECMWF EPS
Skill
Spread
10-members
BSresol
BSrel
10-members
Multi-analysis EPS
• MA EPS: 6-member ensemble
• Compare with EPS for 500 hPa height,
spring 2002 (90 cases)
• Spread less than EPS
• Worse probability scores than EPS
Solid red: EPS; dash blue: MA EPS
Solid red: EPS; dash blue: MA EPS
Possible Revisions to EPS 2003-2004
1. Twice a day running (12z and 0z)
+improved scheduling
2. Dry T42 singular vectors 48hr optimisation
 Moist T63 singular vectors 24hr
optimisation
3. TL255L40TL319-TL399L65
4. Hessian (possibly RRKF) metric
Dry vs moist SVs
27/12/99. M.Coutinho,
Reading U
24-hr optimisation
T63 resolution
Dry vs moist SVs
15/10/87
Dry vs moist SVs
2/8/97
To find the initial perturbation, consistent with
the statistics of initial error, which evolves into
the perturbation with largest total energy
xt0 , M  Mxt0 
xt , xt 
max
 max

1
1
x t  0 x t , A x t 
x t  0



x
t
,
A
xt0 
0
0
0
0
0

M  M xt0    A1xt0 
Singular vectors of M
In principle, A is the analysis error covariance matrix. In practice, A is
approximated by a simplified metric (eg total energy)
M t ,t0 
Isopleth
of initial
pdf
Isopleth of
forecast pdf
Initial time metric and SV structure
Singular vectors for T1/Lothar computed
with different initial time metrics
• total
energy, Hessian metric with/without
observations
•optimization period: 24 Dec 1999, 12 UT +48h
Initial time metric and SV structure
temperature at 45N of leading SV optimized for Europe
Hessian
{
Total energy
Initial time metric and SV structure
Vertical correlations 700hPa, 5leading SVs optimized for Europe
Total energy
Let X the state vector in an NWP model
X  F[ X ]  P[ X ; ]  R
Terms retained in the
Galerkin basis
projection of the
underlying pde
Local bulk formula
representing the mean
effect of neglected
scales - driven by
resolved scales (eg
diffusion)
Residual, =0 in
most GCMs.
Represent as
stochastic noise
=P in ECMWF
model where  is
a stochastic
variable?
ECMWF stochastic physics scheme(s)
X  D  P   P
i
 is a stochastic variable, drawn from a
uniform distribution in [-0.5, 0.5],
constant over time intervals of 6hrs and
over 10x10 lat/long boxes
X  D  P   ( P  D)
ii
X  D  P   D
iii
2-day forecasts differing only in realisations of
the stochastic physics parametrisation
Stochastic Physics has a positive impact on ensemble skill
Area under ROC curve. E: precip>40mm/day.
Winter- top curves. Summer – bottom curves
Stoch phys
No stoch phys
Buizza et al, 1999