The Data Assimilation System for the Nowcasting Demonstration Project (NDP)
Very short-range forecasting over time periods of just a few hours ahead is commonly referred to as "nowcasting" and accurate observations of the initial conditions are vital.
Since March 2012 the Met Office has been running a demonstration trial of an hourly cycling Numerical Weather Prediction (NWP) system for southern UK. This combines a 1.5 km resolution version of the Met Office Numerical Weather Prediction models and 3 km resolution Data Assimilation Methods. This system is referred to as the Nowcasting Demonstration Project (NDP) and every hour produces analyses and forecasts for a period of 0 to 12 hours.
The NDP has been enabled by the installation of a new supercomputer allowing faster production of more technically advanced NWP forecasts and data assimilation in addition to the current operational forecast systems and means that the NDP is available to Met Office forecasters over the period of the 2012 London Olympic and Paralympic Games. This period allows us to perform a demonstration and evaluation of its capabilities and potential for future operational use over the whole UK if more computer power is available in the future.
The NDP data assimilation system
The extra computer power, and smaller domain of the NDP compared to the The UKV model - Kilometre-scale forecasting over the UK with the Unified Model, has enabled the NDP to use the more computationally expensive 4D-Var data assimilation method and to update the analysis every hour compared to the cheaper 3D-Var used every 3 hours for the The UKV model - Kilometre-scale forecasting over the UK with the Unified Model forecast system.
|Model||Resolution||DA method||DA time window||Cycling||Forecast length|
|UKV (UK)||1.5 km||3D-Var (3 km)||3 h||3 h||T+36 (every 6 h)|
|NDP (Southern UK)||1.5 km||4D-Var (3 km)||1 h||1 h||T+12 (every 1 h)|
4D-Var enables use of high time-frequency sub-hourly observations in the NDP system. 4D-Var ensures the better use of observations away from the nominal analysis time and also allows use of observations at the same location at more than one time in the analysis time window.
The NDP 4D-Var data assimilation is configured to assimilate:
- Doppler radial wind (DRW) observations from 5 radars, 6 times per hour,
- wind from wind profilers every 15 minutes,
- satellite radiances from Meteosat Second Generation (MSG) SEVIRI channel 5 (clear and over low cloud) and channel 6 (clear) plus, over sea only, clear window channels every 15 minutes,
- hourly 3D moisture derived from cloud observations (satellite + surface reports),
- MSG cloud- and humidity-tracked winds (air motion vectors) once an hour, aircraft temperature and winds (AMDAR) once per hour, and
- hourly surface temperature, relative humidity, wind, pressure and visibility.
In addition, radar-derived surface rain rates, available every 15 minutes, are assimilated, in the first hour of the forecast, using the technique of latent-heat nudging - adjustments to the temperature and moisture fields aloft in response to the precipitation.
The model domain (shown in the figure above) was chosen to include the 5 Doppler radars in southern England that are currently providing radial winds approved as suitable quality for assimilation. Doppler radial winds from Ingham radar were only available from July 2012.
The data assimilation time window runs from 30 minutes before the hour to 30 minutes after. We wait 15 minutes for observations to arrive so start the forecast at 45 minutes after the hour. The time taken by the 4D-Var depends on the number of observations. Since the DRWs are obtained by reflection of radar pulses from rain or ice cloud, there are more of these when it is raining. It may be possible in future to have even more timely forecasts by having an off-centred data assimilation period of 50 minutes before the hour to 10 minutes after the hour.
Observations are not available uniformly in space, nor are all required variables available at all observation locations. Therefore their information needs to be spread smoothly horizontally and vertically to adjust the previous forecast fields and also via so called balance relationships to adjust related fields e.g. pressure producing wind increments and vice versa due to the geostrophic relationship and pressure producing temperature increments and vice versa due to the hydrostatic balance equation. These balance relationships break down at convective scales when non-hydrostatic ageostrophic motion can predominate, especially in severe thunderstorms etc but for the time being they are still used in the data assimilation system where the analysed variables are increments of stream function, velocity potential, ageostrophic pressure, total relative humidity and logarithm of aerosol mixing ratio (logm). This spreading of information is determined by the ratio of observation and forecasts error variances and forecast background error covariances derived from differences between different length forecasts valid for the same time which used the same lateral boundary conditions. For the UKV these are 24 and 12 hour forecast differences and for the NDP they are 6 and 3 hour differences.
The horizontal spreading uses a second-order auto-regressive (SOAR) correlation function to model horizontal correlations. For the UKV these have long length scales derived from observation- forecast differences at radiosonde locations of 180 km for streamfunction, 130 km for velocity potential and unbalanced pressure and 90 km for humidity and logm. For the NDP they have been derived directly from the forecast differences and vary with vertical mode between 60 and 10 km for velocity potential and stream function and between 30 km and 2 km for both unbalanced pressure and humidity. The graphic below shows the changes in the potential temperature field at approximately 850 hPa due to the assimilation of a temperature increment of 2 K with an observation error of 1 K. On the left is an assimilation using the NDP background error covariances and on the right is an assimilation using the UKV background error covariances.
The use of Doppler radial wind was introduced operationally in the UKV model, once every 3 hours, from July 2011. At that time radial winds were not yet available from Ingham radar so only winds from Chenies, Clee Hill, Dean Hill and Cobbacombe Cross radars were included. Prior to use in the UKV they were trialled in the NDP system using hourly 3D-Var with the UKV covariance statistics for 4 cases of approximately 17 cycles each. This showed a benefit of approximately one hour's skill in precipitation forecasts out to 5 hours, see figure at the right showing Fraction Skill Score for rain accumulation exceeding 0.2 mm over 1 hour and a scale of 55 km. The blue line is the skill of the control forecasts without Doppler Radial winds and the red line is the skill of the forecasts including the Doppler radial winds.
The impact of sub-hourly Doppler Radial Winds in NDP 4D-Var was also tested in a couple of cases with 16 and 23 cycles respectively. The figure at the right shows the rms error in m/s of the forecasts compared to the Doppler radar wind observations as a function of forecast range. The black line is for NDP forecasts produced from hourly cycling 3D-Var with Doppler radial wind used once per hour and the green line is for NDP forecasts using 4D-Var with Doppler radial winds 6 times per hour. The benefit of using 4D-Var with sub-hourly data over 3D-Var can be clearly seen.
More Information on the NDP system and example forecasts can be found by following links below:
- The Nowcasting Demonstration Project for the London 2012 Olympic and Paralympic Games for the London 2012 Olympic and Paralympic Games
- The Nowcasting Demonstration Project: case study of 7 May 2012 Oxfordshire tornado Oxfordshire tornado
- The Nowcasting Demonstration Project: case study of line convection on 28 May 2012 thunderstorm line from Kent to Northamptonshire/Leicestershire border