AOSC Departmental Seminar
April 23, 2015

Developing Ensemble-Based Tools to Improve and Understand the Predictability of High-Impact Events


Brian Ancell
Department of Geosciences, Texas Tech University
Abstract:  

Ensemble forecasting provides well-known advantages over deterministic forecasting, primarily in the form of probability/uncertainty and enhanced skill through the ensemble mean.  However, it may be possible to further improve ensemble predictability by extracting information within the ensemble specific to chosen high-impact events.  Two such techniques are 1) choosing the best ensemble members based on their proximity to the mean, and 2) choosing ensemble subsets that have the smallest errors in regions determined by ensemble sensitivity analysis.  The first technique is aimed at providing a realistic forecast (on the model attractor) at times when the ensemble mean, which is typically best on average, diverges from the model attractor and becomes unrealistic in the presence of significant nonlinear perturbation evolution.  It is found that over a large number of midlatitude cyclones, the best forecast is created by patching together members that have the smallest differences from the mean with regard to the cyclones at different forecast times.  This is because individual members become closer and further from the mean in seemingly random ways, even though the ensemble spread of the cyclones grows steadily.  Consequences of such "patched-together" forecasts will be discussed.

The second technique attempts to identify the best ensemble members early in a forecast period by comparing observations to the members in sensitive areas (determined through ensemble sensitivity analysis).  Ensemble subsets are chosen based on this method, and shown to provide improved prediction of midlatitude cyclones over that of the full ensemble.  This initial examination was fairly idealized, and used a perfect model/perfect observation OSSE setup.  Nonetheless, these results motivate further development of the technique for potential realistic forecast benefits.  Since both techniques discussed above have shown success at synoptic scales, it is still a question whether they can show similar success at convective scales that possess higher nonlinearity.  Along these lines, an initial look at convective sensitivity analysis will be shown.  The integration of the proposed techniques into an operational tool at the National Weather Service will be discussed.