Mesoscale Ensemble Data Assimilation and Prediction System

Background

Fig. 1. Comparison of the RMSE of 12hr forecasts of temperatures (T), relative humidity (Rh), and zonal wind component (U) of WRF that are initialized from 12h data assimilation using Cressman-type observation-nudging (FDDA), 4DREKF (REKF), WRFDA 3DVAR and 4 DVAR, DART-EnKF (EAKF), and two versions of GSI (GSIm and GSIn) for an OSSE study. The WRF Forecast without data assimilation is denoted as CTL.
Fig. 1. Comparison of the RMSE of 12hr forecasts of temperatures (T), relative humidity (Rh), and zonal wind component (U) of WRF that are initialized from 12h data assimilation using Cressman-type observation-nudging (FDDA), 4DREKF (REKF), WRFDA 3DVAR and 4 DVAR, DART-EnKF (EAKF), and two versions of GSI (GSIm and GSIn) for an OSSE study. The WRF Forecast without data assimilation is denoted as CTL.

Given the chaotic nature of the atmosphere and the imperfections of numerical weather prediction (NWP) models, probabilistic forecasts are imperative for applications. To address this need, RAL has been developing an Ensemble Real-Time Four-Dimensional Data Assimilation (E-RTFDDA) and forecasting system. This WRF-based mesoscale ensemble has been deployed to support US Army test range operation and real-time wind energy prediction. Unlike most other mesoscale ensemble systems, E-RTFDDA is a multi–model, multi–scale, and rapidly cycling data assimilation and prediction system with multiple perturbation approaches. The continuous cycling mechanism of E-RTFDDA allows the model to produce accurate nowcasts and short-term forecasts. Since 2011, an innovative ensemble data assimilation algorithm known as four-dimensional relaxation ensemble Kalman filter (4D-REKF) has been under development, to replace the simpler Cressman-type “observation-nudging” FDDA in E-RTFDDA with a flow-dependent weighting. The research and development of E-RTFDDA is currently through the sponsorship of the Army Test and Evaluation Command, Xcel Energy and China State Grid projects, whose broad objectives and progress are also discussed in this annual report.

Research and development activities for the Ensemble RTFDDA (E-RTFDDA) during FY2014 were focused in three areas: 1) enhancements to the ensemble perturbation methods, 2) development of 4D-REKF, and 3) ensemble system deployment including advances in post-processing and calibration.

In FY2014, two E-RTFDDA systems, each with 15 WRF and 15 MM5 members, were continuously operated: one at the US Army Dugway Proving Ground to support routine tests and test planning, and one for Xcel Energy to provide ensemble wind forecasts at Xcel facilities located in Colorado, Minnesota, Texas and New Mexico. In addition, two new 30-WRF-member E-RTFDDA systems were implemented to provide wind forecasts to support China State Grid wind integration. Bias correction and probability calibration have been developed to produce value-added probabilistic forecast products for end users.     

RAL continued to enhance the E–RTFDDA perturbations approaches, adding WRF-NMM (non-hydrostatic mesoscale model) members to the system; the Canadian GEM (Global Environment Multiscale) and Japanese Meteorological Administration GSM (Global Spectral atmospheric Model) model outputs were used to enhance E-RTFDDA boundary condition perturbations; the NCAR DART ensemble Kalman filter (EnKF) tools were also integrated with E-RTFDDA to enhance E-RTFDDA initial condition perturbations. The NCAR DART-EnKF (Data Assimilation Research Testbed-Ensemble Kalman Filter) system was integrated into E-RTFDDA to enhance the E-RTFDDA system in both member perturbations and data assimilation. The enhancement allows DART EnKF to take advantage of E-RTFDDA by means of deriving error covariance using the multiple perturbation E-RTFDDA forecasts; meanwhile, the updated EnKF means and a subset of the EnKF members are used to perturb the initial conditions in E-RTFDDA.

Major work on 4D-REKF included testing and optimizing the 4D-REKF code and algorithms, preparing for real-time operational runs. Because it is impractical to directly compute and communicate the entire Kalman gain to the nudging scheme, several more efficient approaches for computing the Kalman gain, while maintaining adequate accuracy, were tested. The schemes include a 2D mathematical fitting function, a ray-tracing method and a "nearest-point" approach. It was found that the "nearest-point" method presents the best trade-off between accuracy and efficiency for the Kalman gain computation. To handle  diverse observation types, two kinds of Kalman gain computation were designed. For fixed stations, the Kalman gains are calculated at the exact station locations. For time-space changing observations such as aircraft weather reports, the Kalman gains are computed at regular grid points and then interpolated to the observation locations.

To validate the 4D-REKF data assimilation system, OSSE experiments based on perfect-model and perfect-observations assumptions were conducted. For comparison, WRF 3DVAR, 4DVAR, Gridpoint Statistical Inerpolation (GSI), and NCAR DART-EnKF have also been tested with the same OSSE framework.  The verification results of this OSSE study indicate that 4D-REKF outperforms other existing WRF data assimilation technologies. Figure 1 is an example showing the bias and RMSE of various meteorological variables in 4D-REKF at the end of 6-hour data assimilation in comparison to those of WRF 3DVAR, 4DVAR, NCAR DART-EnKF, and GSI. The bias and RMSE in the standard FDDA (distance-dependent weighting) are also shown in the figure.

The 4D-REKF FDDA scheme was deployed as a component of the operational E-RTFDDA system running at the US Army Dugway Proving Ground, Utah. Assessment of the impact of 4D-REKF is on going, but the preliminary results suggest that the raw Kalman gains computed from the ensemble forecasts are not sufficiently accurate. Further refinements to the Kalman gains estimation and new hybrid Kalman gain schemes were formulated and are being tested.

FY2015 Plans

E-RTFDDA will be further enhanced in the following areas:

  • Evaluation of WRF-NMM, SKEBS, and DART EnKF components and conduct forecasting experiments to provide guidance for configuring optimal E-RTFDDA systems according to the requirements of special applications.
  • Integration and assessment of more global model outputs to improve the specification of the lateral boundary conditions perturbation for E-RTFDDA. 
  • Continued refinement of ensemble Kalman gain computation for 4D-REKF and enhance 4D-REKF with cross-variable (covariance) observation-nudging capabilities. In particular, this new capability will be applied to assimilate Doppler-radar radial velocities.
  • Implementation of 4D-REKF for real-time operations at US Army Dugway Proving Ground, and verify the advantages.
  • Continued improvement of the ensemble forecast verification and calibration algorithms. The Quantile Regression (QR) calibration approach will be reformulated to train the correction according to weather regimes. Improvement of the definition and determination of analogs for post-processing will also be considered.
  • Continued E-RTFDDA technology transfer through developing new collaborations with U.S. and international agencies.