An Approach to Improving the 4DEnVar-Initialized Deterministic Prediction Skill for Global Weather

Global deterministic numerical weather predictions (DNWPs) initialized from the ensemble mean analysis of four-dimensional ensemble-variational (4DEnVar) data assimilation are usually worse than those initialized by four-dimensional variational (4DVar) data assimilation. As the average of ensemble forecasts initialized from 4DEnVar analyses generally outperforms the 4DVar-initialized DNWP in anomaly correlation and anomaly root mean square error, a new approach was proposed to improve the 4DEnVar-initialized DNWP with only doubling forecasting computational resources. This method approximately supplements the diffusion term existing in the ensemble mean forecast, which is missed in the DNWP initialized from ensemble mean analysis. The diffusion filters noise from initial uncertainty. Experiments show higher skill in the new 4DEnVar-initialized DNWP than in those initialized from 4DEnVar ensemble mean analysis using purely ensemble covariances and 4DVar analysis with a purely climatological covariance. This approach provides an effective way to expand applications of 4DEnVar and possibly other ensemble-based methods to operational forecasting.