ASSIMILATION of RADAR DATA at CONVECTIVE SCALES with the EnKF: PERFECT-MODEL EXPERIMENTS USING WRF / DART Altuğ Aksoy National Center for Atmospheric Research Collaborators: Chris Snyder (NCAR) and David Dowell (CIMMS)
MAIN OBJECTIVES / CHALLENGES OF THE PROJECT Exploration of the EnKF using the WRF / DART system Real radar data for multiple convective cases with differing behavior (supercell, linear, multi-cellular). Some of the challenges / questions: –Can we implement a certain filter configuration to assimilate in situations with varying convective characteristics? –How do we best assimilate reflectivity observations? –How do we best initialize our ensemble to account for the differing characteristics across cases? –How do we deal with the varying number of observations on the radar grid (too many near the radar vs. too little away from the radar)? –How do we deal with mature cells propagating into the domain? (For this project, we will limit ourselves to convective initiation within the domain.)
CURRENT WRF / DART SYSTEM CAPABILITIES System has been improved to simulate and assimilate radar observations (reflectivity and radial velocity) WRF (Weather Research Forecast) model V. 2.1: –Can be run in “idealized” mode –Open boundary conditions –Initial state obtained from a given sounding –Perturbations are obtained through placement of thermal bubbles DART (Data Assimilation Research Testbed) V. iceland: –Offers different ensemble-based schemes –Capable of assimilating radar observations –2D or 3D localization –Covariance inflation in model space –Adaptive covariance inflation in observation space
EXPERIMENT DETAILS - WRF 150 km x 150 km x 18 km at 2-km horizontal resolution –Domain centered at KOUN radar (Norman, Oklahoma) –Flat terrain –WRF standard sounding for the “quarter-circle shear supercell test case” –Perturbations: Control: 5-K thermal bubble near the radar (60 x 60 km) Two initialization methods explored: –First method: 5 thermal bubbles randomly placed in a sub-domain (location selection from uniform distribution) –Second method: 2 thermal bubbles randomly placed around observed cell (Gaussian distribution with standard deviation of 8 km) –In both methods, thermal amplitude is selected from Gaussian (4K, 2K) No perturbations applied to the background environment
CONTROL RUN (60 min FORECAST) 20 Minutes40 Minutes60 Minutes 3-km Reflectivity (dBZ) and Horizontal Winds 7-km Vertical Wind (m/s, colored) and Surface Negative Temperature Perturbation (2K Contours)
EXPERIMENT DETAILS - EnKF Ensemble size: 50 members 2-D localization with influence half width of ~6 km No covariance inflation Ensemble forecast initialized at 10-min control time (when first >10 dBZ reflectivity is observed and the cell is identified) First assimilation performed at 20-min ens. forecast (30- min control) 5 assimilation cycles performed with 5-min ens. forecasts in-between
EXPERIMENT DETAILS - OBSERVATIONS Both reflectivity and radial velocity assimilated Only observations with reflectivity > 10 dBZ assimilated Observations simulated on radar grid. There are critical differences from generating obs at model grid: –Because obs locations do not generally overlap with model grids, the forward operator involves spatial approximation to the model grid –Number of observations assimilated depends on the relative locations of the storm and the radar: When the observed storm is near the radar, one faces the danger of assimilating too many observations, hence filter divergence. When the observed storm is away from the radar, there may not be enough observations to approach the truth. In this case, number of observations (R > 10 dBZ) grows from ~6,000 at the first cycle to ~13,000 at the fifth cycle (mainly due to storm intensification). Observation error (std. dev.): 5 dBZ for reflectivity and 2 m/s for radial velocity
COMPARISON OF GAUSSIAN and UNIFORM INITIAL DISTRIBUTION of BUBBLES (RAINWATER MIXING RATIO, kg/kg) Time (minutes) At Observation Locations Ratio of Variance to Square Mean Error Over Entire Domain RMS Error Ne / (Ne+1)
ERROR and SPREAD GAUSSIAN INITIAL BUBBLE DISTRIBUTION (VERTICAL WIND, m/s) Time (minutes) At Observation Locations Ratio of Variance to Square Mean Error At Model Reflectivity > 10 dBZ RMS Error Ne / (Ne+1)
ERROR and SPREAD GAUSSIAN INITIAL BUBBLE DISTRIBUTION (ZONAL WIND, m/s) Time (minutes) At Observation Locations Ratio of Variance to Square Mean Error At Model Reflectivity > 10 dBZ RMS Error Ne / (Ne+1)
ERROR and SPREAD GAUSSIAN INITIAL BUBBLE DISTRIBUTION (TEMPERATURE, K) Time (minutes) At Observation Locations Ratio of Variance to Square Mean Error At Model Reflectivity > 10 dBZ RMS Error Ne / (Ne+1)
COMPARISON OF ENSEMBLE MEAN TO THE CONTROL FORECAST HOR. DISTRIBUTION OF 7-km VERTICAL WIND (m/s, colored) and SFC COLD POOL (2-K contours) ForecastAnalysisControl (Truth) 30 min Assim. Cycle 1 50 min Assim. Cycle 5
COMPARISON OF ENSEMBLE MEAN TO THE CONTROL FORECAST VERTICAL CROSS SECTION OF TEMPERATURE PERT. (K, colored) and VERTICAL WIND (4 m/s contours) ForecastAnalysisControl (Truth) 30 min Assim. Cycle 1 50 min Assim. Cycle 5
FINAL THOUGHTS Initiation with bubbles with Gaussian displacement is observed to perform better, but may need tuning. Thresholding of reflectivity has been the only successful way to deal with the large number of observations. However, –There is valuable information in 0-dBZ observations we would like to utilize, especially when members have spurious cells –Methods exist to adaptively apply covariance inflation when observation density is high. Could be expensive… –Adaptive assimilation could also be performed depending on the magnitude of innovations