Mesoscale Ensemble Data Assimilation and Prediction


Figure 1: A description of E-RTFDDA framework
Figure 1: A description of E-RTFDDA framework

Given the chaotic nature of the atmosphere and the imperfections of numerical weather prediction (NWP) models, probabilistic forecasts are imperative for mesoscale weather forecasting. In the last 10 years, RAL developed an Ensemble Real-Time Four-Dimensional Data Assimilation (E-RTFDDA) and forecasting system. The system is built upon WRF. The first E-RTFDDA system was deployed to support US Army Dugway Proving Ground in 2007, known as E-4DWX. Since then the WRF core, data assimilation scheme, ensemble perturbation approaches, and ensemble output post-processing have been continuously improved. The second system was developed in 2010 to support Xcel Energy for real-time wind energy prediction. Recently, in working with the Chinese Electric Power Research Institute (CEPRI), the third system, a 30-member 9km-grid E-RTFDDA model that covers China, has been developed. 


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 it to produce accurate nowcasts and short-term forecasts that are highly desirable for many weather-critical applications. One of the unique advantages of the E-RTFDDA system is that it is capable of integrating and dynamically downscaling the top-ranked global model forecasts, including the forecast data from ECMWF/IFS (Europe), NCEP/GFS (USA), EC/GEM (Canada), NASA/GEOS (USA), JMA/GSM(Japan), and CMA/GRAPES (China), to the same mesoscale application grids, and thus effectively takes in the information and uncertainties of these global models and produces more accurate mesoscale probability forecasts. Figure 1 presents a high-level description of the E-RTFDDA technology. 

Another E-RTFDDA core component is its innovative ensemble data-assimilation algorithm, four-dimensional relaxation ensemble Kalman filter (4D-REKF). In the ensemble modeling, 4D-REKF permits flow-dependent data weighting, which improves the simpler Cressman-type “observation-nudging” FDDA. 4D-REKF computes Kalman gains using the multi-model E-RTFDDA forecasts, which are ingested into E-RTFDDA models to replace the simple distance-dependent observation weighting functions in the original nudging model. A Local Ensemble Kalman Filter (LEKF) is employed to take account of multiple observations. 4D-REKF retains and leverages the advantages of both traditional Newtonian-relaxation and Ensemble Kalman Filter data assimilation schemes. It eliminates the shortcoming of empirical specification of spatial weighting functions in the current station-nudging FDDA formulation. Furthermore, it extends the traditional (intermittent) EnKF data assimilation method to a 4D continuous data assimilation paradigm that greatly mitigates the dynamic shocks associated with the intermittent EnKF processes. 4D-REKF reduces the critical dependency on the background error covariance inflation required by the traditional EnKF and permits effective assimilation of all observations that may be available at irregular locations and times.  Figure 2 describes the general formulation of the 4D-REKF FDDA scheme. 4D-REKF enhances both the accuracy of model initial conditions and also the initial condition perturbation approach, and thus improves the overall capabilities of E-RTFDDA ensemble prediction. 

Figure 2: A description of Obs-nudging and 4D-REKF FDDA formulations.
Figure 2: A description of the observation-nudging and 4D-REKF FDDA formulations.

The research and development of E-RTFDDA is currently conducted under the sponsorship of the US Army Test and Evaluation Command, the State Grid Corporation of China (SGCC), Xcel Energy, the State Power Investment of China (SPIC), and Inner Mongolia Electric Power Corporation (IMEPC) for supporting military tests, electric-power-grid weather safety, and wind-power forecasts, respectively. 


The E-RTFDDA WRF model has been upgraded to the newest community WRF release Version 3.9.1 and 4.0. The upgrades involved evaluating and adopting the community WRF advances, implementation of an analog-based bias-correction algorithm and a quantile-regression-based probability-calibration scheme for statistical post-processing. The Army E-4DWX system was redesigned to have a 3-km domain for four test ranges and one week of test runs were completed.  An E-RTFDDA system was used to support the power grid operation of the State Grid Corporation of China (SGCC) and wind power forecasts for three wind farms by the State Power Investment of China (SPIC). The SGCC E-RTFDDA system contains 30 WRF members and its domain cover the whole China region at grid intervals of 9 km. The system runs four forecast cycles per day, and each cycle produces 72-hour forecasts.  For the SPIC wind power forecasting, a 30-member 3-km E-RTFDDA system is developed for the SPIC wind farms over a region with complex terrain. The system runs 3-hour cycles, producing 93-hour forecasts.  

Besides continuous R&D of the existing operational ensemble system, a new ensemble system has been developed for the Inner Mongolia Electric Power Corporation (IMEPC) for the purpose of forecasting wind power in a very-large-scale wind power production region. Figure 3 shows the domain configuration of the modeling system. The wind farm clusters are modeled with two 2.7-km domains. The system contains 45 ensemble members.

Figure 3. The domain configuration of the modeling system
Figure 3. The domain configuration of the modeling system


The WRF stochastic kinematic energy backscattering (SKEB) scheme for dynamical ensemble perturbation and DART-EnKF for WRF initial condition perturbation strategies will be further evaluated and integrated for real-time E-RTFDDA operation.

4D-REKF is still in its early stages of initial operating capability. Further evaluation and enhancement are necessary to fully exploit the power of the technology.  The limited representativeness of Kalman gains computed from mesoscale ensemble forecasts of excessively small number of ensemble members is still the main challenge for effectively taking the advantage of 4D-REKF.  Empirical algorithms will be explored to address the fact that ensemble mesoscale forecasts often lead to formation of sporadic, unrepresentative local structures in the Kalman gains, which introduces noise and lessens the effectiveness of data assimilation. Refinement to 4D-REKF with cross-variable (covariance) “observation-nudging” capabilities for assimilating Doppler radar radial velocities should be studied. Strategies for nudging hydrometeors (rain, snow, and graupel mixing ratios) and radar reflectivity and lightning observations for cloud-resolvable ensemble forecast of convective systems will be studied. Finally, 4D-REKF for assimilating wind farm data to improve wind-power forecasts will be conducted in 2020.