Data assimilation for nowcastingpotential and limits of a 3D variational approach

How to use the newer, better NWP models to help nowcasting applications

● the AROME system: status and plans● 3DVar vs other techniques● The concepts of balance & control● Redefining the NWC/NWP boundary

Example of kilometric-scale NWP model: AROME

● a new mesoscale convection-permitting NWP system built from ECMWF's IFS, Europe's ALADIN, and France's Meso-NH models

● Efficient spectral, semi-Lagrangian, semi-implicit NH compressible numerics to allow fast real-time production

● Reasonably sophisticated physics: prognostic TKE turbulence, 5-species prognostic cloud microphysics, RRTM/FM radiation, tiled surface scheme with soil, vegetation, lakes, sea, snow, town energy balance, high-quality physiographies

● With own data assimilation using radar, satellite, in situ operational observations

● 1-way nesting in 10-km ALADIN data assimilation, itself nested in 20-km ARPEGE global 4DVar assimilation

Impact of NWP model resolution: 10km vs 2.5km, fields of low-level wind

(blue) and T (red) on the model grids

(different wind scaling in each figure)

Arome MCS simulation (04-08-94 15 to 18 UTC)2,5 km / dt=15s / domain 144 * 144 / analysis Diagpack + Humidity bogus

Arome-2.5km 9h-range fog dissipation forecast

Meteosat visible image

The AROME data assimilation● derived from ECMWF 4D-Var, plus mesoscale features● 3DVar algorithm with FGAT (first guess at appropriate time)

allowing 1-min time resolution, with 1-hour cycling● Multivariate non-separable Jb structure functions derived from

ensemble statistics● Variational relaxation of large scales to coupling model● Use of automated screen-level obs network (T, Td, wind) with

variational control of PBL stability ● Direct multivariate assimilation of geostationary IR radiances in

clear air (control of tropospheric humidity)● (planned) 1D cloud bogussing, starting with nowcasts of convective

clouds (ISIS/RDT software)● (planned) Direct multivariate assimilation of radial Doppler winds

from radars, and 1D radar precipitation bogussing

Dyn. Adapt. Raingauges

3DVar 3DVar with SEVIRI

Impact study :Precipitation forecast

2004/07/18 12UTCRR P12 – P6

Objective score impact of 10km assimilation vs. range (rmse and bias)

(pink=ARPEGE 4DVar dynamical adpation, blue=ALADIN LAM 3DVar)

RH2mRR

AROME real-time forecast on 21 June 2005

AROME fc

dynamical adaptation

AROME fc started

from mesoscale assimilation

15 TU

radar composite

AROME status & plans

● 2.5km forecast model runs daily since May 2005 on 500km domain with 1-minute timestep

● Excellent performance on wind, low-level temperature and convective weather

● Quality is situation-dependent: long routine verification is needed● Assimilation runs at lower 10km resolution so far with very positive

impact on 0-12h forecast ranges wrt. dynamical adaptation● main target: 6-hourly 36-h NWP forecasts over France

(1000kmx1000km) in less than 30 minutes, in 2008 + hourly very short-range forecasts

● priority on relocatable nowcasting applications in 2009-2010● see presentations by G Jaubert, V Ducrocq, O Caumont, T Bergot

3DVar vs other techniques● 3DVar is complex software, but numerically cheap i.e. quick

(unpreconditioned ALADIN 3DVar converges in 50 iterations i.e. about 5 minutes)

● 4DVar would take at least 10 times more computing, delaying forecasts by tens of minutes: serious handicap for short-range NWP

● short-window 4DVar works well for Doppler wind processing● Kalman filter can beat 3DVar in theory without the timeliness

penalty (heavy computations are done out of the critical path) but not as mature yet for operations

● 3DVar physical foundation makes it nicely extensible to new observation types (e.g. the ever-changing satellites)

● future algorithm: probably a 3DVar basis mixed with short-window 4DVar + an ensemble KF focused on sensitive phenomena

Usable observations for convective systems assimilation

Concepts of balance & control● 3DVar smoothing functions & multivariate relationships must be specified a

priori by a « background Jb term » forecast error model:– either you have observations of the phenomena that drive the

prediction: e.g. PBL humidity and convergence lines for convection initiation --- the choice of DA algorithm will not matter

– or, you have indirect observations and you need to spatialize them using likely Jb multivariate structures: local, weather-dependent balance properties, to retrieve the driving phenomenon

● It is often better to observe causes than effects (e.g.: ground precip)● Automatic model feedbacks & static Jbs work better for large scales

(geostrophism, Ekman pumping...) than mesoscales (PBL tops, orography, 3D convective & frontal structures)

● Two competing strategies at mesoscale:– « automatic » balance estimation: 4DVar and (ensemble)KF– « ad hoc » spatialization: object bogussing from image processing

Perspective:From NWP to Nowcasting

● challenge 1: refresh NWP forecasts faster than forecast error growth– will require ad hoc structuring of NWP production systems (Rapid

Update Cycle, decentralized computing or superfast telecoms)● challenge 2: produce short-term direct forecasts of observables and end

user products– simulation of satellite, radar etc. output at high resolution & quality– human monitoring tool to intercept/correct poor model output

● challenge 3: intelligent probabilistic post-processing of hard-to-model weather elements e.g. storm risk areas vs. exact Cb cell location

How can we help humans to cope with increasing data volumes

of irregular quality ?