1. Workshop on Use of NWP for Nowcasting October 2011
Development of NWP-based Nowcasting at the Met Office -The Nowcasting Demonstration Project
Workshop on Use of NWP for Nowcasting October 2011
Susan P Ballard, Zhihong Li, David Simonin, Jean-Francois Caron, Cristina Charlton-Perez, Nicolas Gaussiat, Lee Hawkness-Smith, Helen Buttery, Graeme Kelly, Robert Tubbs, Dingmin Li +DAE Exeter + Humphrey Lean, Emilie Carter, Nigel Roberts….
© Crown copyright Met Office

2. Contents Introduction and Plans
NWP based Nowcasting compared to UKPP conventional nowcasts
Impact of Doppler Winds and Observation Errors
Impact of Observations and Background Errors
Impact of Analysis Increments and Boundary Conditions
Conclusions and future work
© Crown copyright Met Office

3. Aims Hourly cycling 1.5km resolution NWP for southern England - NDP
6-12hour forecasts (12hours would allow 6 member lagged ensembles)
Needs new supercomputer
Olympics demonstration
Compare UKV (UK 1.5km 3hourly 3D-Var), UKPP (conventional nowcast) and NDP (nowcasting demonstration project)

4. Nowcasting Model Domain
Resolution
VAR
Time Window
Cycling
Forecast Length
UK4 / UKV
4 km / 1.5km
3D-Var 4/3km
3 hr
T+36
South UK Fixed
1.5 km
3D/4D-Var 1.5/3km
1 hr
T+6 or T+12
The Nowcasting model domain at present can be nested (1-way) in either one of 2 operational limited area models.
This means that the Lateral Boundary Conditions (LBCs) are given by either
UK4 model (4 km) or
UKV model (1.5 km variable resolution with grid stretched from 1.5 to 4 km at the boundary)
© Crown copyright Met Office 4

5. Nowcasting Data Assimilation Strategy
Techniques:
3DVAR, 4DVAR, latent heat and moisture nudging
Needs:
Hourly analyses and forecasts to customer within 15mins of data time
Therefore data must be in Met Office in real time
With 4D-Var can exploit high spatial and temporal resolution
High temporal resolution eg every 5 -15mins may help to offset poor or limited horizontal resolution
Exploit more observations – type and time frequency eg
GPS, AMDAR, Meteosat imagery (clear and cloudy)
use of radar Doppler winds, reflectivity and refractivity data

6. UKPP – Met Office Nowcast
– forecasts of surface rain rate valid at 21Z 3/6/2007
T+3
T+2
radar
T+1
T+0
© Crown copyright Met Office

7. Impact of hourly 4D-Var data assimilation
Including latent heat nudging of 15min
Radar derived rain rates
– forecasts of surface rain rate valid at 21Z 3/6/2007
T+3
T+2
radar
T+1
T+0
Zhihong Li
© Crown copyright Met Office

8. Impact of Doppler Radial Winds on Fractional Skill Score for rainfall >0.2mm accumulation – scale 55km
Forecast skill is improved for rain accumulations with low threshold – effectively all rain (~ 1 hour gain)
FSS – Roberts and Lean
ΔFSS
ΔFSS – 0.2 mm acc – scale 55km
© Crown copyright Met Office 8

9. Impact of Doppler Radial Wind Assimilation on RMSE differences in fit to surface winds and Satwinds
RepError= error derived from O-B statistics approx 2-3m/s
Obs Error = standard deviation of superobs approx 1-2m/s
© Crown copyright Met Office 9

10. Doppler Wind Observation Error
4 months of data O-B
Increase with range
Range from 2.4 to 3.2 m/s
© Crown copyright Met Office 10

11. Hollingsworth–Lonnberg
Observation Error
Hollingsworth–Lonnberg
Rely on the use of departure between the background and observation (innovations)
Construct a histogram of background departure covariance against distance
Distance
σ2o
σ2b
Background error covariance
I included this slide to show how high resolution data, in addition to being quality controlled must be super-obbed or “summarised” in a sense.
Because the model grid is so much coarser than the observation grid, the observations will not be independent pieces of information as far as the model is concerned.
We have to give the model a meaningful ob, and to do this we have to be intelligent about how we condense the observational information.
There are multiple ways to condense the obs, so you have to test your design against others to be sure you have a good metho
With this approach fewer raw observations influence the super-observation near the radar than at longer measurement ranges.

12. 15Z 18th May 2011
UKV T+6
radar
UKV T+0

13. Impact of Observations T+0 NDP
All obs
No obs
No doppler
wind
No GPS
Zhihong Li

14. Background Errors Jean-Francois Caron
NMC method – Met Office - SOAR horizontal correlation function
UKV T+24/T+12, 30 cases
180km for streamfunction, 130km for velocity potential and unbalanced pressure and 90km for humidity and logm
UKV T+6/T+3
130 to 30km for velocity potential and stream function
60 to 5km for unbalanced pressure and 40 to 5km for humidity
NDP T+6/T+3 every 6 hours, 75
60 -10km vel pot, stream fn , 30-2km unbalanced pressure and 30-2km for humidity
Ensembles – Meteo France
Brousseau et al 2011
6 members, 26 days, 3hour range

15. Impact of new Cov Stats and MOPS/LHN in NDP T+0
All obs
Zhihong Li/Jean-Francois Caron
radar

16. Impact of background errors on screen Temperature

17. Balance and Boundary Updates - rms pstar tendency - NDP
Zhihong Li
4D-Var Analysis
increments
at T-30mins,
start of window T-30mins
Top – 15Z, 9Z lbc
Middle – 16Z, 15Z lbc
Bottom – 17Z, 15Z lbc

18. Conclusions
Operational convective scale NWP beating advection type precipitation nowcast from about T+2.5hours
Good progress being made in use of radar data in NWP-based nowcasting
However many challenges still to extract the full benefit from the radar data and other observations
Only just getting access to observations to test real benefit in NWP-based nowcasting
Modifications to background errors, linear model and other options in DA likely to improve skill
© Crown copyright Met Office 18

19. Other Scientific Challenges
Matching the observations during the assimilation cycle and first 2 hours of forecast
Balance and control variables, adaptive vertical grid
Precipitation bias in rates and area – data assimilation and modelling problems. Model too cellular and can miss some convection – maybe need for more 3dimensional parametrizations?
Computer resources – should be able to produce forecast within 1 hour 19

20. Questions and answers
© Crown copyright Met Office 20

21. 6 March 2011 15UTC – Robert Tubbs
SEVIRI Channel 5
UKV Analysis simulated Ch 5
SEVIRI Channel 9
UKV Analysis simulated Ch 9

22. Thunderstorms 28th June 2011 12UTC
UKV T+3
UK4 T+3
radar

23. Thunderstorms 28th June 2011 12UTC
UKV T+9
UK4 T+3
radar

24. Radar Rain Rates 28/6/2011
9UTC
10UTC
12UTC
11UTC

25. Hourly cycling NDP – 12UTC original version
Zhihong Li
T+3
T+2
T+1
T+0

26. NDP forecasts for 12UTC 28 June 2011 latest UKV-type version
Zhihong Li
T+2
T+3
T+0
T+1

27. UKPP forecasts for 12UTC 28 June 2011

28. Background Aim: to replace current UKPP nowcasting system
Hazardous weather , especially flood risk
Boscastle, North Cornwall
16th August 2004
Estimated Cost £500million
Data Assimilation/analysis vital
for these short period forecasts
of 0-6 hours

29. Current Nowcasts - UKPP
UKPP analysis of surface rain rate every 5 mins at 2km
Radar composite plus 2DVAR of UK4 and MSG outside radar area
Nowcasts every 15mins to 7 hours using T-30, T-15 and T+0 rain analyses to derive field of motion
Blend UK4 and nowcast using STEPS
8 member ensemble
Hourly temperature, precip type, cloud, wind, visibility etc
Start 1min after DT but waits 7mins for radar rates and satellite imagery
Available within 15mins 29

30. Comparison of NWP forecasts and STEPS (advection) Nowcasts of Precipitation – RMSF error of 1hour accumulation > 1mm
RMSF=
Exp(Sum/N)0.5
Sum= sum(i-1,N)
Log(Fi/Oi)2
where
O=radar estimate
Both smoothed to
6km 30

31. Nowcasting System
NWP-based nowcasting system is non-hydrostatic model that resolves convection explicitly.
Nonlinear Unified Model (UM) is 1.5 km resolution (360 x 288 x 70 levels), has model top at 40 km, and uses 50 s timestep.
Linear Perturbation Forecast (PF) model and its adjoint: 3 km resolution (180 x 144 x 70 levels) and 100 s timestep.
4D-Var uses hourly assimilation windows with linearization states for PF model updated every 10 minutes and UKV every 10mins in OPS
Observations are extracted in the observation time window T-30 to T+30 minutes. Might change to T-60 to T+0.
4D-Var increments are added to UM at initial forecast time T-30 mins (first UM time step). Might change to T-60.
Runs using latest UKV as boundary conditions, 45min cutoff time
OPS creates CX files with hourly data for the UK4/UKV but 10 minute intervals for the NDP.
GCR Generalised Conjugate Residual solver has a tolerance set too low for the NDP. GCR algorithm solves Ax=b for x where A is non-symmetric.
Zhihong “As with mesoscale, global N320L70 and MOGREPS-G N216L70 configurations, it has been noted that there are large oscillations in surface pressure (p1) in SUKF nowcasting system. These noises are due to the UM's GCR solver not converging with sufficient accuracy. Currently, GCR_TOL_ABS is set to 1.0e-3 with GCR_RESTART_VALUE of 2 in the SUKF UM GCR solver. These results show that 1.0e-4 for GCR_TOL_ABS is the best setting without a large increase in the CPU time (by about 5% compared with the original setting of 1.0e-3). GCR_TOL_ABS=1.0e-4 is therefore recommended for the SUKF nowcasting system if we can afford doing so. GCR_TOL_ABS=2.5e-4 could be a good compromise as it increases CPU time by only about 1.5% with most noises originally present disappearing.”

32. Hourly 3D/4D-Var DA cycle
Assimilation/Obs window T-0.5 to T+0.5
LH and cloud (RH) nudging from T-1 to T+0
Schematic diagram of 3D/4D-Var DA cycle with MOPS cloud and LH nudging in SUKF 1.5km hourly nowcasting system

33. Observations currently available
Surface fields (u, v, p ,T, RH, visibility) [1 min] 1 hr
Wind profiler (u, v) [15 min] 15min
Doppler radar radial winds [5 min] 3 per hr NDP*
Radar reflectivity both direct and indirect [5 min] not used
Rain rates (radar-derived) [15 min] 15 min NDP
GPS (integrated humidity path) [10 min] 10min
Scatterometer winds [~5 min] 1 hr
Cloud-track winds (Atmospheric Motion Vectors AMVs)
[low res. 30 min, high res. 15 min] low res. only used 1 hr
VAD Velocity Azimuthal Display winds [1 hr] 1 hr NDP*
Radiosondes [12 hours or when reported] when reported NDP
3D cloud analysis (MOPS) [1 hr] 1 hr NDP
AMDAR aircraft-derived T, u, v [1 hour batches] 1 hr NDP
Geostationary Satellite Imagery (clear radiances) [15 min] 1 hr NDP
Dark blue =time that the data is available. Light blue =time data is used.
* on VAD winds NDP means that VAD winds are not used if we are using the Doppler radial winds
John Jones runs GPS AC (Analysis Centre), known as METO; We also get data from French AC, SGN and the German AC, GFZ.
John is working on a new GPS processing system which will process data in 15-minute batches, rather than the current hourly solution.
Clear-sky radiances: over land only the upper tropospheric water channels are used. Over the sea, all IR channels are used over the vertical.
Helen B.: Doppler radial wind measurements are currently available at six radar stations in the UK (out to a maximum radius of 100 km at various elevations (between 1 and 9 degrees)).
Observations:every 1 degree azimuthally; every 600 m radially, every 5 minutes.
Hope to improve local representation of wind velocity and location of convergence/divergence and hence of local rain features.
Some of the doppler data will be assimilated operationally in UK 4 and UKV systems shortly. Currently the UK4 and UKV assimilate VAD winds derived from the same data as the Doppler radial winds, but provide an average wind velocity that varies with height at only one horizontal location, that of the radar station.
Trials carried out in the UK4 using 3D Var with Doppler radial winds assimilated at the four operational Doppler radar stations in the South of England; the control assimilated VAD winds at those stations instead. Result: The UK Index showed a neutral impact but benefit was seen in the location of the precipitation when assessed with the fraction skill score

34. Observations coming soon
Potential for assimilation of following observations:
Cloud track winds (u, v) or Atmospheric Motion Vectors (AMVs) at high resolution available but not yet assimilated operationally
Radar reflectivity (indirect and direct assimilation)
Satellite-derived cloud top
Surface cloud cover and cloud base (manual and automatic reporting)
Lower resolution fields of cloud track winds are already assimilated in the UK4 and UKV models. We haven’t put them in the Nowcasting.
Sat cloud top and surface cloud will be assimilated directly soon, but these data are already put into MOPS the 3D cloud analysis package which is assimilated.
Radar reflectivity can be assim. directly into 4DVar or indirectly by first using the reflectivity to translate to temperature and specific humidity profiles. From Nicolas’ document: “Measured radar 3D reflectivities that result from consecutive scans at different elevation show the time evolution of clouds and precipitation at a very hi resolution. They contain crucial dynamic and thermodynamic information as precipitation development is linked to changes in horizontal convergence and moisture fields. If radar reflectivities can be assimilated using 4D-Var, it should be possible to extract this dynamical and thermodynamical information directly. Using 3D-Var require the use of an optimal inversion technique prior to the assimilation in order to link cloud and precipitation information to a realistic increment in moisture and dynamical fields. “
Ceilometer cloud base recorder data is already going into the MOPS but is not directly assimilated.