Ensemble filter data assimilation: Application to probabilistic nowcasting of boundary layer profile

1
Ensemble filter data assimilation Application
to probabilistic nowcasting of boundary layer
profiles

• Dorita Rostkier-EdelsteinIIBR, ILJosh P.
  HackerNCAR, CO

2
Outline

• Ensemble data assimilation (DA)
• Planetary boundary layer (PBL) forecasts
• Assimilation of surface observations
• Single column model (SCM)
• Numerical experiments
• Summary and on-going work

3
DA in geophysics

• The use of all the available information and its
  uncertainty in order to estimate as accurate as
  possible the present state of the atmospheric or
  oceanic flow.
• The available information consists of present
  information provided by observations and of past
  information provided by model dynamical
  calculations and past observations.
• (based on definitions by Talagrand and
  Ghil)
• There is also a synergistic effect between
  observations and dynamics (firstly observed by
  Thompson)
• The model dynamics propagate the information
  contained in the observations to areas where
  observations are sparse.

4
The DA equation
5
Background error covariance
Observations error covariance
Analysis error covariance
6
Background error covariance

• How are these calculated?
• Two examples
• 3DVar, previous talk
• From climatological runs.
• Flow independent at assimilation time.
• Ensemble filter DA
• From an ensemble of runs at real time.
• Flow dependent at assimilation time.

7
Ensemble filter DA

• An ensemble of N data assimilation cycles is
  carried out simultaneously.
• All the cycles assimilate the same real
  observations, different sets of random
  perturbations are added to the observations or to
  the forecasts (background).
• Get and ensemble of N analysis and forecasts at
  time ti-1, estimate the forecast error covariance
  from the N forecasts
• Provides an estimate of the flow dependent
  uncertainty in the analysis/forecast.

2 for 1 Improved initial conditions forecast
uncertainty
8
Moores Law Calculations per second per 1K
9
The Planetary Boundary Layer (PBL)

• The bottom of the troposphere that is in contact
  with the surface of the Earth.
• Transient coupling and decoupling with the
  Earths surface and the free atmosphere aloft
• Short scale dynamics need to be parameterized in
  mesoscale numerical weather prediction (NWP)
  models.
• Model error in the PBL as severe as other sources
  of error.
• Successful forecast of the PBL is important for
  convective initiation and forecasted
  precipitation, air pollution and transport,
  frontal propagation, sea breeze, slope flows,
  visibility

10
Surface observations

• They are often the most reliable and easiest to
  set up observations we have in the PBL
• They are generally under-utilized in NWP and DA
• Difficult to determine the vertical influence of
  the observations.
• Error of representativeness (coarse resolution of
  the model).
• Dynamic balances exploited in large-scale DA
  inappropriate.
• Potential for information transfer from the
  surface to the atmosphere aloft.

Intermittent, anisotropic, nonstationary
Correlation between the variable at the surface
and in the column above. WRF Model Climatology,
4km resolution, Oklahoma, May 3-July 14, 2003
11
Bring together

• Ensemble DA
• Surface observations
• Forecast the state of the PBL with a simple model

12
Single column model (SCM) of the PBL

• Simulates the atmosphere and earth surface in a
  single grid column and is forced by observational
  or model data.
• Internal dynamics for ageostrophic wind,
  diffusion eq., etc.
• Geostrophic and radiative forcing from a
  mesoscale 3D-model (e.g., WRF) or observations.
• Full suite of WRF-ARW land surface (LS), surface
  layer (SL) and PBL parameterizations (plus
  additional options).
• In general proved useful for nowcasting (0-12
  hours).
• Fast, then possible to run many ensemble
  realizations.
• Original model development by Mariusz Pagowski,
  NOAA/ESRL.

13
PBL state estimation with an ensemble filter,
surface observations and a SCM First experiments

• Goal Prove the usefulness of ensemble
  assimilation of surface observations in
  determining the state of the PBL.
• Location and period Oklahoma, May 3-July 14,
  2003.
• Assimilated variables Wind (at 10 m height),
  Temperature (at 2 m height), Mixing ratio (at 2 m
  height).
• First guess Profiles extracted and perturbed
  from WRF climatology.
• Profile estimation is based on
• climatological first guess and real time surface
  information
• ONLY

14
First experiments, OKlahoma Verification against
rawinsondes 00Z, 19 LT, MAE
Meteorological information from Climatological
profiles Surface observations ONLY
15
An estimate of the flow dependent uncertainty is
also provided
Ensemble spread, single realizations
Zonal wind component, U (m/s)
Potential temperature (?C)
Mixing ratio (g/kg)
A calibrated ensemble will provide a measure of
the flow dependent uncertainty
16
Improved system Factor separation analysis

• Goal Find an efficient method for probabilistic
  nowcasting of PBL profiles.
• Improved system
• First guess is centered on WRF forecasts valid at
  time of surface observations (dated first guess
  instead of climatology), scaled climatological
  perturbations.
• SCM includes now
• Horizontal advection from WRF forecasts,
  dynamically tuned by assimilation (state
  augmentation).
• Parameterized radiation (in particular for
  nighttime radiative cooling).
• Complexity has a cost in both computation and
  maintenance.
• Investigate the importance of 3 factors
• Surface assimilation
• Horizontal advection
• Parameterized radiation.

17
Factor separation analysis

• First introduced into meteorological research by
  Stein and Alpert, 1993.
• Application to the present system
• Run 23 combinations of factors assimilation,
  advection, parameterized radiation.
• Verification of each combination.
• Single factors are easy to asses, but synergistic
  contributions to verification metrics can be
  difficult.

F.S.
18
Effect of factors on Mean Absolute Error (MAE)
Potential temperature
Day time
Night time
(A negative contribution reduces MAE, only most
significant factors are shown)

• Assimilation only has significant effect in
  reducing error.
• (weak horizontal advection at convective hours)

• Assimilation most significant.
• Advection and radiation play roles at different
  heights.
• Assimilation-advection synergism increases error
  as it partially destroys the assimilation benefit.

19
Effect of factors on Mean Absolute Error (MAE)
U-wind component
Day time
Night time
(A negative contribution reduces MAE, only most
significant factors are shown)

• Assimilation and advection comparable effect in
  the PBL.
• Advection most important aloft.
• Synergism is again increasing error

20
Profile nowcasting at Beit-Dagan

• First guess WRF climatology or WRF forecasts at
  horizontal resolution of 10 km (no significant
  differences with WRF at finer resolutions, i.e.,
  3.3 and 1.1 km grid size)
• Assimilation of surface observations.
• Verification against radiosondes.

00 UTC July 2006
21
Summary and on-going work

• The results show a big impact of assimilating
  surface observations with an ensemble filter on
  the skill of estimated PBL profiles under
  different weather regimes (e.g., US continental
  plains, IL coastal area).
• Factor separation leads to a rapid interpretation
  of DA in the context of other factors.
• Advection has often a positive effect as it
  brings 3D dynamics.
• But advection can be harmful when the coupling is
  strong.
• Radiation plays a role at night.
• Present on-going work focuses on
• Factor separation applied to probabilistic
  verification of the forecast ensemble.
• Upgrade of the SCM to include the full WRF
  capabilities (i.e., vertical velocity) WRF-SCM.
• Future work retrieval of PBL heights.

22
WRF and SCM, Oklahoma
23
(No Transcript)