Real Time WRF EnKF Data Assimilation and Forecast System


matlab graphic

matlab graphic


matlab graphic

matlab graphic


Verification graphics not implemented yet. 04/01/2011


36-km EnKF
Starting (right) going back 8 cycles (left). matlab graphic

4-km EnKF
Starting (right) going back 8 cycles (left). matlab graphic



An Ensemble Kalman filter (EnKF) is a data assimilation technique for making atmospheric analyses and forecasts. Data assimilation is the process whereby information from model forecasts and observations are combined to give the "best" estimate of the state of the atmosphere. Although the EnKF is relatively new, it shows great potential for scale-independent data assimilation, because it relaxes assumptions that are made in current operational schemes. Furthermore, it provides a sample of equally likely analyses that may be used for ensemble forecasts without the need for generating perturbations (i.e. singular vectors and bred grown modes are not needed).

A key aspect of the data assimilation process involves the error statistics for the observations and for the model forecast. Although observation error statistics are fairly well known, error statistics for the model forecast are typically assumed, despite the fact that they change continually. These error statistics are important because they provide the relative weighting given to observations and model forecasts in making an analysis. For example, if observation errors are small compared to the model forecast, then more weight is given to the observation. In addition to weighting the observations relative to the model, the statistics also provide the information needed to spread observational information to other locations, and other variables. For example, a pressure observation that increase the pressure in the analysis at a point should be expected to also produce an anticyclonic adjustment to the analysis wind field around the point, if the flow is "balanced."

The EnKF generates flow-dependent statistics of model forecast errors by creating an ensemble of forecasts; here we use a 64 member ensemble. One analysis is produced for each ensemble member by estimating the forecast-error statistics from the ensemble (we assume Gaussian statistics, so the ensemble mean and covariance matrix provide the necessary statistics). Given the 64 member analysis ensemble, 64 new forecasts can be created by running the model again; the cycle is repeated indefinitely.


The real-time system along with web graphics and design has been developed by Philip Regulski at the University of Washington. Rahul Mahajan has been involved with major contributions and wrote the DART driver code. Prof. Greg Hakim is the principal investigator and with Prof. Cliff Mass they supervise the development of the system and direct the project. David Warren and Dr. Harry Edmon built the linux computer cluster on which this system runs, and provide continual systems administrative support and expert guidance. Neal Johnson provides data and backup support.

Details of the UW-Dept. of Atmospheric Sciences implementation:

The system uses The Data Assimilation Research Testbed (DART) for the data assimilation and the model forecasts use the Weather Research and Forecasting model (WRF). DART is a software environment that provides a variety of data assimilation methods developed and maintained by the Data Assimilation Research Section (DAReS) at the National Center for Atmospheric Research (NCAR). The system uses the WRF model (v3.0.1), run at 36 km and 4km horizontal resolution. It uses the WSM 3-class simple ice microphysics, RRTM longwave radiation, Dudhia shortwave radiation, Monin-Obukhov surface-layer, YSU boundary-layer (PBL) and the Kain-Fritsch cumulus schemes along with the Noah land-surface model. Currently the system has 64 members. Observations are assimilated every 3 hours (00, 03, 06, 09, 12, 15, 18 and 21 UTC) and 24-hr forecasts are started every 6 hours (00, 06, 12 and 18 UTC).

Observations include a dense local surface network along with upperair observations. Assimilated data consists of radiosondes on mandatory levels (u, v, T, RH), ACARS data (u, v, T), cloud track winds (u, v), fixed buoy pressure, drifting buoy pressure and surface station data (alt, u, v, T) for stations where the difference between model topography and station elevation is less than 300 meters. All the observations go through multiple quality control assurance tests by two independently developed QC systems. A typical cycle assimilates between 8,000-15,000 observations.


How to change images:

  1. Select an initialization time.
  2. Select a domain (36- or 4-km resolution)
  3. Select a plot type.
  4. Chose your plot.
  5. Press Plot button.

Browser issues:

For best viewing results please expand your browser to fill the screen. Smaller screens may not be able to view the page as intended (the right image may wrap below).

This page has been tested using IE9, Firefox4, Chrome 10.0+, Safari 5.0+ and Opera 11.0+. For best results please upgrade your browsers to at least these or the newest versions.

Please direct web questions and comments to: David Ovens

Validate XHTML Valid CSS!