Bridging high-resolution model and coarse ensemble system: Hybrid Ensembling

Both high-resolution single deterministic model and reduced-resolution ensemble model are currently run side-by-side at major NWP centers such as NCEP and ECMWF. Due to limited computing resources, these two tasks are sometimes viewed as competitors to each other. As a matter of fact, each task has its own shortness and merit in reality. Generally speaking, hi-res single run is more accurate and more detailed in structures but provides no uncertainty information at all, while coarse ensemble system does provide uncertainty information but is less accurate and lack of detailed spatial structures. Therefore, it's necessary to consider how to combine them into one unified system to retain both accuracy, detailed spatial structure and forecast uncertainty information.

In April 2004, a method called "Hybrid Ensembling" was proposed and tested with the NCEP Short Range Ensemble Forecasting (SREF) system (Du 2004, available at http://wwwt.emc.ncep.noaa.gov/mmb/SREF/reference.html). This method essentially applies the spread or uncertainty information from the coarse ensemble system to the forecast with the most "detail" and least overall error. The 2004 paper evaluates mainly the overall improvement in forecast performance measured by various scores, while this one will focus on the improvement in detailed spatial structures.

A potential problem of this method is that coarse ensemble and hi-res run might drift apart from each other too far. Our preliminary experiments show that it's not likely to be a problem for short-range (0-3 days) although it could be a problem for medium- or long-range (say after a week, but need to be investigated). To prevent this dilemma, an interesting question could be researched is that what's the best ratio of horizontal resolutions between hi-res single model and coarse ensemble?