Title: Numerical weather prediction with high local resolution
1M. Tsyrulnikov and E. Astakhova
Hydrometcenter of Russia Current Status and
Progress of Data Assimilation and Ensemble
Forecasting
NCEP/EMC, May 15, 2007
2Outline
- Operational technology
- A new 3D-Var scheme under construction
- First results with the new analysis scheme in
the ocean - Current research in satellite data assimilation
- First results in ensemble prediction
- Model-error stochastic modelling (a simulation
study)
3Operational data assimilation (M Tsyrulnikov, A
Bagrov and R Zaripov)
- 3-D multivariate OI based analysis (to be
replaced by 3D-Var) - Observations TEMP, SYNOP, SHIP, BUOY, SATEM,
Cloud-motion and Aircraft winds, SHIP and SATOB
sea surface temperatures - Intermittent assimilation
- - global with the semi-Lagrangian or the
spectral model of the HMC - - regional/meso experiments with WRF, plans
with LM - - static analyses with UKMO or NCEP 6-h.
forecasts as the first guess
4The Global Spectral Model
- Resolution T85L31 ( 1.41x 1.41?) -
Semi-implicit time scheme, time step900 s -
Non-linear normal mode initialisation - MPI
parallelization Operational runs twice a day
0000 UTC 84-h forecast 1200 UTC 240-h
forecast
5The spectral model Physics
- Large-scale condensation and precipitation
scheme, including - evaporation of rain
- Â Â Kuo-type convection scheme
- Â Â Surface processes Land surface temperature,
soil moisture - and snow
depth are predicted -
SST is fixed - Â Â Â Â Vertical diffusion Surface layer
Monin-Obukhov theory -
Free atmosphere fluxes depend on the
-
mixing length and the Richardson number - Â Â Â Â Radiation scheme Mean path length method
-
H2O, CO2, O3, aerosol -
5 spectral intervals -
Diurnal cycle -
Albedo depends on the snow height
6The global semi-Lagrangian model(M Tolstykh)
- Â Resolution 1.125/1.40625? lat/lon,
- 28 s- levels
- Model domain Globe, option with rotated pole
- 4th order compact differences for discretization
of derivatives in non-advective terms, including
semi-implicit scheme and U-V reconstruction - Physics from Meteo-France
- MPI OpenMP parallelization
- Quasi-operational runs
- 0000 UTC 84-h forecast 1200 UTC 240-h
forecast
7RosHydromet modernization project (started in
2007) Financing
Source USD, Millions
International Bank for Reconstruction and Development 80
Borrower Government of Russia 52.5
Total 132.5
8Project Structure
- Technical re-equipment of computation and
communications systems - Modernization of observation networks
- Enhancement of the institutional structure,
improvement of information delivery techniques
and emergency preparedness - Hydrometcenter of Russia a 8-Tfl computer with
1000 PEs
9Development of the new 3D-Var scheme (M
Tsyrulnikov and P Svirenko)
Requirements (1) Universality. The new scheme is
being built to be applicable on global, regional,
and meso scale in the atmosphere and also for
oceanic data assimilation (2) Applicability in a
future EnKF data assimilation system (in
particular, capability of representing spatially
variable flow dependent structures)
10The new covariance model SARMA
The general design
Cf. the NCEPs approach
ARMA in the vertical
The model implemented
11Ref. Tsyroulnikov M.D. Proportionality of
scales an isotropy-like property of geophysical
fields. - Quart. J. Roy. Meteorol. Soc., 2001, v.
127, N 578, 2741-2760.
12The new covariance model operators
The stochastic model
The horizontal operators
or
13Ref. Tsyroulnikov M.D. Proportionality of
scales an isotropy-like property of geophysical
fields. - Quart. J. Roy. Meteorol. Soc., 2001, v.
127, N 578, 2741-2760.
14The analysis formulation
In the computation of the gain matrix
15The new covariance model implementation
- Compactly supported covariance generating
functions - Sparse numerical linear algebra is used
- - The resulting univariate 3D-Var analysis scheme
takes 5-10 min. on a single Itanium processor - - Horizontal isotropy is built in, but can be
relaxed - - Full degree of non-separability in 3-D
correlations - - Spatial variability of the field structure
16Vertically variable horizontal length scale (a
realization of the SARMA pseudo-random field)
17Vertically variable vertical length scale
18Horizontally variable horizontal length scale
In the spot at the center of the plot, the
horizontal length scale is specified to be 1/3 of
that over the rest of the globe
19Locally tilted structures The vertical
cross-section of a realization of the random field
20Local horizontal anisotropy. Single-obs. test
results
Isotropic case
Anisotropy 45 deg. from N
21M.Tsyrulnikov and P.Svirenko
Review of the new 3D-Var system
- The new covariance model based on 3-D spatial
filters is proposed, tested, and implemented in a
3D-Var system - The model is formulated
constructively and so is guaranteed to be well
posed - The model is coordinate-free and so can
be used for data assimilation in the atmosphere
or in the ocean on any domain and in any
geometry - The model is capable of representing
realistic covariances with full degree of 3-D
non-separability - The capability to model
various spatially variable field structures is
demonstrated
22Oceanic data assimilation (A Zelenko, Yu
Resnyansky, M Tsyrulnikov, B Strukov and P
Svirenko)
- A primitive-equations oceanic forecast model
(global, 2 deg. resolution) (similar to GFDL
z-coordinate model with rigid lid at the
surface) - Upper ocean mixed layer (integral) model
similar to that of Kraus Turner (1967) with
updated parameterization of TKE budget and
special algorithm for imbedding into OGCM
(Resnyansky Zelenko, 1991) - - A simplified 2D-Var SARMA analysis scheme
- Operational observations (GTS messages BUOY,
BATHY, TESAC Argo floats transmitted in TESAC
being the most influential) - Runs operationally since August 2006
23Salinity analyses (surface layer)
MERCATOR
HMC 2D-Var
24Analyses of the speed of currents (surface layer)
MERCATOR
HMC 2D-Var
25Satellite data assimilation research first
attempts
- - GPS data assimilation (implementation of the
research by M. Gorbunov) - - A study aimed at the estimation of the
likelihood function for scatterometer wind
observations is underway - An investigation into spatial correlations for
microwave satellite radiance observations is
planned - Current plans (1) AMSU-A and (2) GPS
26The observation operator(M Gorbunov)
1. The Canonical Transform technique for GPS data
processing based on Fourier Integral Operators
was proposed by Ref M. E. Gorbunov and K. B.
Lauritsen. Analysis of wave fields by Fourier
Integral Operators and its application for radio
occultations, Radio Science, 39(4), RS4010,
10.1029/2003RS002971, 2004. 2. The observational
operator relies on the Canonical Transform
technique and provides high accuracy and
vertical resolution of 50-100 m when retrieving
bending angle profiles. Ref M. E. Gorbunov and
L. Kornblueh. Principles of Variational
Assimilation of GNSS Radio Occultation Data,
Report No. 350, Max Planck Institute for
Meteorology, Hamburg, December, 2003.
27Occultation 0041, 2004.01.18UTC 0409, 26.0S
19.4E (by M Gorbunov)
28Occultation 0097, 2004.01.18UTC 0950, 78.8N
125.6W
29Summary of the data assimilation R/D (M
Tsyrulnikov, P Svirenko, R Zaripov, M Gorbunov, A
Zelenko, Yu Resnyansky and B Strukov)
- The unified SARMA covariance model is proposed
and implemented in the analysis system - The 3-D univariate atmospheric analysis OK, the
multivariate analysis under construction - The oceanic 2D-Var is operational since 2006
- The satellite assimilation is at an early stage
30Ensemble forecasting with the breeding method
first attempts (E Astakhova)
-
- Based on T85L31 model
- Breeding with global rescaling and the total
energy norm - Breeding step 12 h
- Analysis increment as an initial perturbation
- 9 members
- Implemented on a 2-processor (Xeon, 1.6 GHz)
server
31516 552 576 contours at 500 hPa. 120h forecast.
Dec. 1, 2006
32Meso-scale multi-model ensemble (N Veltishev)
- MM5
- WRF
- Eta
- LM
- Two domestic models (regional and meso)
- GTS transmitted ECMWF, NCEP, UKMO, and DWD
forecasts - emphasis on near-surface fields
33Ensemble forecasts by system simulation
System simulation (a) simulate analysis
errors (b) simulate model tendency errors Need
for adequate stochastic models
34Stochastic modelling of model errors (M
Tsyrulnikov)
A simulation study truth shallow-water
model model vorticity equation 1. Critical
importance of state-dependent model errors
2. The general design of the stochastic model
35Stochastic modelling of model errors
Model errors appear to be advected, with the
effective advection velocity very different from
the wind
36Stochastic modelling of model errors
In ensemble prediction experiments, the
identified and estimated model-error model had
almost perfect performance
37Stochastic modelling of model errors practical
conclusions of the simulation study
When modelling model errors (1) Look for a
locally (in physical space) state-dependent model
errors component (state-dependence of the first
kind) (2) The residual (after subtraction of
state-dependent errors) model errors can be
regarded as stochastic, satisfying a linear
advection-dissipation equation forced by the
white-in-time noise (3) Parameters of this
equation (variance and local spatial
length-scales of the driving noise, the effective
advection velocity vector, and the dissipation
time scale) are, most likely, also dependent on
the local flow structure (state-dependence of the
second kind)
38Stochastic modelling of model errors
- Practical stochastic modelling is planned for
2008-2009 - Implementation in an ensemble
prediction system (2010) - Implementation in an
EnKF data assimilation system (2010-2011) Ref. Ts
yrulnikov M.D. Stochastic modelling of model
errors A simulation study. - Quart. J. Roy.
Meteorol. Soc., 2005, v. 131, 3345-3371 .
39Mikhail Tsyrulnikov and Elena Astakhova
Conclusions
- The data assimilation and ensemble forecasting
research and development at the HMC will be
carried out along the following lines - The 3D-Var analysis scheme (2007-2008)
- The Ensemble Prediction System based on system
simulation (2009-2010) - The Ensemble Kalman Filter based data
assimilation system (2009-2011)
40Areas of possible joint research
- Flow-dependent covariance modelling
- Satellite data assimilation (specification of
obs-error covariances) - Ensemble prediction
- Model-error modelling