Development of a Hybrid 4DEnVar and Ensemble Kalman Filter Analysis and Forecast System for Convective scale NWP

Research for data assimilation (3/4DVAR, EnKF and Hybrid methods) is reaching a considerable state of maturity within the meteorological data assimilation (DA) community in the context of large-scale and synoptic-scale flows. However, it is often difficult to simply extend these techniques to nonhydrostatic flows at convective-scale. In recent years, a realtime, weather adaptive hybrid ensemble variational analysis and forecast system with the WRF-ARW as forecast model has been developed for the NOAA supported Warn-on-Forecast project (WoF). The goal is to provide ensemble-based physically consistent gridded analysis and forecast products to forecasters for making timely warning decisions. The analysis domain is determined based on current weather situation of a day. Both the WRF-DART with 36 ensemble members and the 3DEnVar system incorporate available mesoscale forecasts, radar data, satellite retrieved cloud water path, and traditional observations to perform two separate 15-minute data assimilation cycles. The ensemble covariance derived from the 36 ensemble members of the cycled WRF-DART forecasts is used in the 3DEnVAR system. Then 18 ensemble members and one deterministic forecast with 3 hour forecast length are launched every 30 minutes from these high frequency data assimilation cycles. The above system has been tested during the 2017 and 2018 HWT spring experiment period, but only in convection-allowing mode with 3 km horizontal resolution due to computational intensiveness. Recently we have extended the original system to a hybrid 4DEnVar and WRF-DART EnKF system. First, the time dimension is added by using three time levels of the background fields and observations in the variational system, so it is renamed as 4DEnVar. Second, a time-expanded sampling approach (Xu et al. 2008, Mon. Wea. Rev.).is developed in the WRF-DART analysis system. By doing so, the number of required ensemble model predictions, or the ensemble size can be significantly smaller. We expect that the computational cost of the new approach will be much cheaper than the original hybrid system, but maintains the accuracy of the original hybrid analysis thanks to the application of the temporal information in the analysis. The potential merits of this system will be demonstrated by two real data cases and the results will be reported in the conference.