1. Development of convective-scale data assimilation techniques for 0-12h high impact weather forecasting
JuanzhenSun
NCAR, Boulder, Colorado
Oct 25, 2011

2. Outline Introduction Success and Issues Future challenges
- Unique aspects of convective-scale DA
- Overview of techniques
Success and Issues
Future challenges
This talk is in the context of
Warm-season QPF
Radar observations
NCAR experiences
Oct 25, 2011

3. What makes convective-scale DA different?
Objective
- QPF, high-impact weather nowcast/forecast
- Forecast accuracy: county/city scale
Predictability of high-impact weather systems
- Rapid error growth
- Small-scale with multiple scale interaction
Observations
- Limited high-resolution in-situ observations
- Remote sensing:
high resolution, but limited coverage, limited and indirect
variables

4. Convective-scale DA strategies
Place storms at right locations
- Warm Start: Cloud analysis, latent heating insertion, saturation adjustment, updraft profiling
Use frequent update
- Sub-hourly; min window for 4DVAR
- Take advantage of high temporal frequency obs.
- Forced by predictability limitation
Consider cloud-scale balance
- Temporal derivative terms should not be neglected
- Different balance from the large-scale
Use different error statistics
- Large-scale error statistics is not applicable
- Research is still lacking

5. Overview of techniques
Techniques based on reflectivity or precipitation
- DFI, nudging, cloud analysis
- Simple and efficient
- No or limited multivariant balance
3D techniques assimilating both RV and RF from radar
- 3DVAR
- Efficient
- Balance is mostly large-scale
4D techniques assimilating both RV and RF
- 4DVAR, EnKF (and its variants)
- Computationally expensive
- Full model balance, but compromised in practice (limited ensemble members, limited assimilation window)

6. Latent Heat Nudging Mei Xu Hydrometeor increment per Δt
ingest radar reflectivity observations (converted to QR/QS)
add tendency terms to model variables QR/QS and T based on the model state and observations
result in thermodynamic and microphysical adjustment
Hydrometeor increment per Δt
(dQR/dt)obs * Δt if QRmod<QRobs&(dQR/dt)obs>0
ΔQR = g *(QRobs- QRmod) if QRmod>QRobs
0 otherwise
Temperature increment
ΔT = CLS/CPM * ΔQR
where CLS is the latent heat of condensation (or fusion)
CPM=CP*(1.+0.8*QV) is the specific heat for moist air

7. Impact of radar data LHN
case
analysis
observation
no radar
with radar LHN

8. Impact of radar data LHN
case
1 h forecast
observation
no radar
with radar LHN

9. Impact of radar data LHN
case
2 h forecast
observation
no radar
with radar LHN

10. Impact of radar data LHN
case
3 h forecast
observation
no radar
with radar LHN

11. Skills for June 11-17, 2009 Front Range Domain
FSS Evaluation
Averaged over 24 forecasts
Analysis period

12. WRF 3DVAR Radar DA Hongli Wang Cost function
For radar DA
Reflectivity data assimilation
- Assimilate rainwater
- Cloud analysis (optional)
- Assimilate saturation water vapor within cloud (optional)
Control variables
- stream function
- unbalanced velocity potential
- unbalanced temperature
- unbalanced surface pressure
- pseudo relative humidity

13. 6-h Forecasts after four 3DVAR cycles
IHOP one-week runs
6-h Forecasts after four 3DVAR cycles
NORD: Control with no radar DA
RV: Assimilate radial velocity
RF: Assimilate reflectivity
RVRF: Assimilate both
One-week FSS skill (5mm) RF
Both RV
Cycled 3DVAR

14. Beijing Results Shuiyong Fan 2-hour forecasts OBS No Radar RV RF
NORD: Control with no radar DA
RV: Assimilate radial velocity
RF: Assimilate reflectivity
RVRF: Assimilate both
FSS skill for four 2009 summer cases
OBS
No Radar RV RF RV RF

15. Diurnal variation of Radar DA impact
Radar DA has longer
positive impact for late
evening initializations
The positive impact
only lasted 4 hours for
morning initializations
It suggests
that the radar DA works
more effectively for
growing storms than
dissipation storms
Dashed lines:
Warm start
Solid lines:
Cold start
00Z
12Z

16. An example of failed forecast
Cold start analysis
An example of failed forecast
Cold start
3DVAR cycled analysis
Cycled 3DVAR RF

17. Can 3DVAR retrieve the tangential wind?
Radial component
Tangential component
Radars with overlap
Corr: 0.724
Analysis
Single radars
Corr: 0.402
Truth
From Sugimoto et al. (2009)

18. Study of a supercell storm using a 4DVAR system VDRAS Sun (2004)
Rainwater correlation
Radial velocity only
Color contour: qr
Observation w qv
RVand RF
Observation
Forecast
RV only
RF only
Reflectivity only qv w
Without radial velocity, the rain falls
out quickly.
Radial velocity assimilation results
in slantwise updraft and moisture, but
not the reflectivity assimilation
Assimilating both RV and RF
consistently outperforms RV or
RF only

19. 4DVAR systems: VDRAS and WRF 4DVAR
Developed for a cloud model
Trajectory is modeled by the nonlinear model
Full adjoint of the cloud model is used to calculate the
gradient in the minimization
Control variables are model prognostic variables
WRF 4DVAR
Developed for WRF model
Trajectory is modeled by the tangent linear model of WRF with reduced physics
Adjoint of the reduced tangent linear model
Control variables follow those in WRF 3DVAR

20. Inserting VDRAS analysis into WRF inner domain
VDRAS 3km
19 UTC 15 June 2002
WRF 9 km

21. 2-h WRF forecasts
valid at
Observation (061302)
No VDRAS
With VDRAS

22. No VDRAS With VDRAS 5-h WRF forecasts valid at 061302
Observation(061305)
With VDRAS

23. WRF 4DVAR Radar Data Assimilation 4-hour forecasts from a case study (13 June 2002)
OBS
3DVAR
4D_RV
4D_RF

24. ETS of 0-6 hour forecast
4D_RF
4D_RV
1 mm
3DVAR
5 mm

25. WRF/DART EnKF Convective scale data assimilation
Glen Romine
Observations
00Z
01Z
02Z
Control (with no storm-scale info) shows convective bloom early.Assim case blooms as well, but resembles obsmoreso than the control, imo.
Mem 1 w/ radar assim
Mem 1 control

26. Future Challenges
DA for nowcasting application requires different configurations
- Frequent updating
- Radar DA crucial for minimizing spinup time
- Different background error statistics
- Multiple pass for observations with different resolutions
- Different DA schemes
- Make better use of surface observations
- Different physics options?
Rapid cycling with/without radar DA can have negative impact
on convective initiation
- Will more frequent updating with radar DA help?
- Diurnal variation of radar DA impact
- The impact also depends on convection type

27. Opportunities and Challenges
Radar DA still a great challenge
- Reflectivity assimilation
> Improve the accuracy of the latent heating and relative
humidity specification in the simple techniques
> Balance with dynamics
> Error statistics
- Radial velocity assimilation
> Retrieval of the tangential component in 3DVAR
> Clear air returns
> Balance with thermodynamics and microphysics

28. Opportunities and Challenges
Challenges for the 4D techniques
- Computation cost
- Large resource required for developing a full 4DVAR
- Choice of control variables for the convective scale in 4DVAR
- Sample issues and maintenance of ensemble
spread for EnKF
- Model errors

29. VDRAS radar data assimilation reveals how cold pools trigger storms 0611 2046 UTC - 0612 1250 UTC
Pert. Temp. (color)
Shear vector (black arrow)
Wind vector at km (brown arrow)
Contour (35 dBZ reflectivity)