Fusing observation- and model-based probability forecasts for the short-term predictions of convection

A number of research studies indicate that the recent history of convection contains a great deal of information on its future state out to 6 hours and beyond (e.g., Golding 2000). The scale of the convective feature often determines the length of time that it will persist and thus be predictable through extrapolation alone (Wilson et al. 1998). Additional work has been done in developing heuristic models of convection that can be used to predict the Lagrangian evolution (growth and decay) of convection (Hand 1996; Pierce 2000; Megenhardt 2005). The National Convective Weather Forecast (NCWF-2) combines extrapolation and a heuristic approach to produce 0-2 hr probabilistic forecasts of convection that are available in real-time for use by the aviation community. Mueller et al. (this conference) describe how this system has been extended out to 6 hours and discuss its performance. While these observation-based techniques perform well in the very short term (e.g., 0-3 hr), their skill decreases rapidly with lead. On the other hand, the skill of numerical weather prediction models is poor initially due to model spin-up issues, but increases with time. Weygandt and Benjamin (2004) mitigate the poor skill of NWP models at short lead times by developing a probabilistic approach to forecasting convection by using an ensemble of time-lagged Rapid Update Cycle (RUC) model forecasts.

The goal of this study is optimally blend these RUC-based convective probability forecasts with 0-6 hr NCWF convective probability forecasts described by Muller et al. to take advantage of the lead-time dependent relative skill of each method. Data from the spring and summer of 2005 are used to develop a summary of verification statistics for RUC-based and NCWF 1-6 hr probabilistic forecasts. The CSI, POD, FAR, and Bias are calculated as a function of lead time and probability level in order to “calibrate” the two systems prior to merger of the forecast fields. These performance parameters are also used to develop a weighting functions that vary with lead time and time of day. The RUC-based and NCWF 0-6 hr probabilistic forecasts are then blended using this weighting function to produce a single merged probabilistic forecast. The new merged forecasts are then evaluated using a few unique cases from the spring/summer of 2005 with results being intercompared with the individual components of the system. Physical explanations for deficiencies and triumphs of the new system (to be called NCWF-6) are given.