Probabilistic Forecasting of High-Shear, Low-CAPE Severe Environments in the Southeastern United States

Previous studies have found that High-Shear, Low-Cape (HSLC) environments are prevalent in the Southeastern United States (SEUS) especially during the cool season and overnight hours, and produce a large proportion of the severe (i.e., tornado and significant wind speed) reports that occur within this region. Current forecasting techniques have difficulty predicting severe events that occur in HSLC conditions due to many reasons including the rapid destabilization of the environment prior to severe convection.

Currently, one of the biggest challenges to National Weather Service (NWS) forecasters is limiting false alarms (i.e., false positives) and raising probability of detection (i.e., POD) of these severe events. It is possible that some of the current widely-used severe weather parameters show skill partly because they are highly correlated with a more relevant parameter or physical explanation, or contain highly correlated parameters in the calculation of a severe composite parameter. Highly correlated parameters provide much of the same information in a predictive equation, and each additional parameter (i.e., predictor) in the equation can introduce more error in the prediction. To alleviate this problem, a statistical procedure was created to analyze hundreds of meteorological parameters and choose a small set of weakly correlated predictors that represent the most predictive information available within the dataset. This procedure was used on a dataset, including official Storm Prediction Center storm reports labeled as event cases and unproductive watch/warning area centroids labeled as null cases, to develop multiple probabilistic models that predict the occurrence of tornadoes and/or significant wind speeds in HSLC conditions. Preliminary results with this method show a decrease in false alarms rates (i.e., reported false alarms were predicted as nulls).

The probabilistic models produced were run in real-time using widely-available forecast data as input to test the operational forecasting utility of the models. Probabilistic models shown are produced using North American Regional Reanalysis (NARR) data, and these models are run in real-time using North American Mesoscale (NAM) forecast data. The NARR and NAM models share similar characteristics, so it is assumed the training dataset used to create the probabilistic models have similar characteristics as the gridded forecast data used as input into the probabilistic models. Research is funded through a CSTAR grant involving research-to-operations collaboration with the NWS Weather Forecasting Offices within SEUS.