Weather and Climate Systems

Improving Global and Hurricane Forecasts Using Time-lagged Ensembles
in the GFS 4DEnVar Hybrid Data Assimilation System

Bo Huang

School of Meteorology

04 May 2016, 3:00 PM

National Weather Center, Room 5600
120 David L. Boren Blvd.
University of Oklahoma
Norman, OK

Four dimensional ensemble-variational hybrid data assimilation system (4DEnVar) based on the gridpoint statistical interpolation (GSI) system was developed and is under pre-implementation test for the Global Forecast System (GFS) at the National Center for Environment Prediction (NCEP). In 4DEnVar, the ensemble error covariance can provide the four-dimensional flow-dependent error covariance estimate avoiding the need of using the tangent linear and adjoin of the forecast model. Experiments using real data have shown a significant improvement of the 4DEnVar over the operational GFS 3DEnVar system (Wang and Lei, 2014).

Sampling error, characterized by the distant spurious correlation or spurious correlation of different variables at the same location, is one of the major concerns in estimating the background error covariance, due to the use of a finite number of ensemble members. The filter will diverge from the truth state if the sampling error is not treated appropriately. Covariance localization is commonly applied to deal with sampling errors by removing the distant spurious correlations. Such distance-dependent localization is demonstrated to be efficient in reducing noisiness and results in an improved error covariance estimate (Hamill et al., 2001). But the inherent limitation of localization is that it may eliminate the distant but realistic correlations that can be simulated by an extremely large-sized ensemble (Miyoshi et al., 2014). Significant increase of ensemble size, however, can be very difficult for the operational model because of the limited computational resources. Hence, this work aims at exploring the use of a time-lagged ensemble method to increase the ensemble size in the GFS 4DEnVar system while incurring minimum computational cost.

In the study, the lagged ensembles were constructed from (1) the freely-available Global Ensemble Forecast System (GEFS) ensemble which has 20 members and is initialized every 6 hours; and (2) the 80-member ensemble from the current 4DEnVar system extended with a longer forecast lead time. The lagged ensemble forecasts will be ingested for 4DEnVar update along with the regular non-lagged 80-member forecasts. As an initial attempt, the ensemble size in the lagged experiment was increased to 320 (referred to as lag320) and it was compared with the baseline 80-member experiment (referred to as ens80). Estimate of computational cost shows that a large amount of cost saving in lag320 comes from the EnKF update by comparing with directly increasing the ensemble size to 320. One single observation experiment shows that lag320 can more effectively reduce sampling errors than ens80. Results from three-week cycling experiments also demonstrated an improvement of lag320 over ens80, in terms of 6-hour temperature and wind forecast. Verification on longer forecasts and on hurricane track forecasts is ongoing and will be presented in the future