Transcript Слайд 1 - rssi.ru

Numerical weather
prediction: current state and
perspectives
M.A.Tolstykh
Institute of Numerical Mathematics
RAS, and
Hydrometcentre of Russia
What is the global atmospheric
model
Atmospheric equations ~ averaged NavierStokes equations on the rotating sphere.
Processes on unresolved scales are
parameterized. Currently, numerical
solution of the equations for resolved
dynamics accounts for ~30 % of total
computations time, the rest is for
parameterizations
Main ways to increase an accuracy of
numerical weather prediction
1) Increasing the horizontal and vertical
resolution of atmospheric models
Requires masssively parallel computations
=> development of new dynamical cores
(new governing equations, new numerical
techniques)
2) Development of new parameterizations
of subgrid-scale processes
3) Improvement of initial conditions
RMS error of 3-day H500 forecast
Current state of global NWP models
• Typical horizontal resolution at the end of
2009 – 20-30 km
• Japan is the leader with 20 km, next year
ECMWF will be the leader with 15 km
The increase of the processor number
necessary for operational implementation of
the SL-AV model
•
•
•
•
70 km, 28 levels – 4 processors
37 km, 50 levels – 40 processors
20 km, 50 levels - about 350 processors
10 km , 100 levels – supposedly 6000
processors
Development of new dynamical
cores for global NWP models
• Currently, a half of global NWP models us
based on spectral techniques
• It scales up to~0.5N_harm* N_openmp(*N_lev)
processors, ~5000 for Т1279.
Forecast Centre
2009
2010
2011
2012
2013
2014
TL799 L91
TL1279 L91
TL1279 L140
TL1279 L140
tbd
tbd
25 km L70
25 km L70
20 km L90
tbd
tbd
tbd
T799c2.4 L70
T799c2.4 L70
T1240c2.4 L90
tbd
tbd
tbd
30 km L60
30 km L60
15 km L70
15 km L70
tbd
tbd
HMC
T169 L31;
T169 L31;
T339 L63;
(Russia)
tbd
tbd
0.37°x0.45° L50
0.19˚x0.225˚L60
tbd
0.72°x0.9° L50
NCEP
T382 L64 (7.5)
T878 L91 (7.5)
25 km L90
25 km L90
25 km L90
25 km L90
(USA)
T190 L64 (16)
T574 L91 (16)
(0.45°x0.3°) L80
(0.45°x0.3°) L80
(0.45x0.3°) L80
(0.45°x0.3°) L80
(0.3°x0.2°) L90
(0.3°x0.3°) L90
20 km L96
20 km L96
20 km L96
10 km L96
10 km L128
tbd
TL959 L60
TL959 L60
TL959 L60
tbd
tbd
tbd
CMA
TL639 L60
TL639 L60
TL639 L60
(China)
GRAPES
GRAPES
GRAPES
GRAPES
GRAPES
25 km L31
26 km L31
(Country)
ECMWF
(Europe)
Met Office
(UK)
Météo France
(France)
DWD
(Germany)
CMC
(Canada)
CPTEC/INPE
(Brazil)
JMA
(Japan)
tbd
50 km L31
KMA
(Korea)
BoM
50 km L31
50 km L31
T426 L40
40 km L50
25 km L70
25 km L70
25 km L90
tbd
ACCESS ~80 km
L50
~40 km L50
25 km L70
25 km L90
tbd
tbd
New dynamical cores of
atmospheric models
• High parallel efficiency, locality of data
• A grid on the sphere with quasiconstant
resolution
• Computational efficiency of numerical
algorithm (sufficiently long time-step)
• Nonhydrostatic formulation (includes
sound waves)
Choice of the grid
• Traditional lat-lon grids have condensed meridians near
the poles (from presentation by W.Skamarock, NCAR)
Evolution of ps, day 9 (Jablonowski test)
BQ (GISS)
GEOS-FV
HOMME
CAM-EUL
CAM-FV-isen
GEOS-FVCUBE
GME
ICON
with =0°, resolution ≈ 1°1°L26
OLAM
hPa
Reduced latitude-longitude grid
• Routinely used in models based on
spectral approach. It is possible to use it in
finite-difference/finite volume models with
specific formulation
• Advantages
- High-order approximations are easily
possible
- Easy to code and parallelize
Shallow-water model
2,4
Test 7 a (1.5 deg)
L2 err x 10-3
2,0
1,6
1,2
7a Full
0,8
7a Red
0,4
7a Red
New
0,0
24
48
72
Lead time
96
120
Developments in parameterizations
of subgrid-scale processes
• Parameterizations depend on horizontal
resolution (examples: deep convection,
microphysics)
• Taking into account exchanges with
adjacent horizontal grid cells (currenly,
most of parameterizations are 1D in
vertical)
New and advanced parameterizations
of subgrid-scale processes
• Advanced land surface parameterization
accounting for hydrology, evolution of
snow cover, freezing/melting, bogs, …
• Deep convection parameterization for
partially resolved case
• Explicit description of microphysical
processes in clouds
• Lake parameterizations
• Boundary layer parameterizations in the
case of strongly stable stratification
Land surface parameterization
«Tile» approach (subcells describing water,
low and high vegetation, etc)
New directions:
• Soil hydrology taking into account adjacent
grid cells
• Biogeochemistry (carbon cycle, dynamical
leaf area index …)
H-TESSEL surface parameterization
scheme (ECMWF)
• The revised hydrology includes spatial variability related to
topography (runoff) and soil texture (drainage)
Slide 19
ECMWF: New microphysics
parameterization
Current Cloud Scheme
New Cloud Scheme
WATER
VAPOUR
CLOUD
FRACTION
CLOUD
Liquid/Ice
Evaporation
CLOUD
FRACTION
PRECIP
Rain/Snow
Slide 20
Impact of initial data on model forecasts
Среднекв. ошибка H500 за март-апрель 2009 г. Северное полушарие
40
35
30
25
20
15
10
5
0
40
35
30
25
20
15
10
5
0
NCEP PLAV (oper) PLAV (ncep) DWD
24
48
H500 S1 (Северное полушарие) за март-апрель 2009 г.
NCEP PLAV (oper) PLAV (ncep) DWD
24
72
48
заблаговременность T850 RMS (Северное полушарие) за март-апрель 2009 г.
3,5
3
2,5
2
1,5
1
0,5
0
NCEP PLAV (oper) PLAV (ncep) DWD
24
заблаговременность
48
72
заблаговременность
72
Data assimilation
• Weight optimally observations and shortrange forecast from previous initial
conditions to create initial conditions for
the model
• Current approaches: 4D-Var and
ensemble Kalman filter
Some directions of development for the
global semi-Lagrangian model SL-AV
• Increasing the scalability of the code from
~300 to 5000 processors
• Replacement of 3D solvers by divide-and
conquer algorithms
• Nonhydrostatic dynamical core
• More advanced land surface parameterization
(bogs, carbon cycle, multilayer soil, soil
hydrology…)
Conclusions
• Challenges of the nearest decade –
development and implementation of global
atmospheric models with the horizontal
resolution 1-10 km.
• New approaches to develop new dynamical
cores and parameterizations
• This requires efficient parallel
implementation on ~ 10000 processors
=============================
We shorten the distance with leading
centres in the field of global NWP
Thank you for attention!