Quantifying the Value of Convection-Allowing Global-Nested Ensembles for Extended-Range Severe Weather Forecasting using Random Forests

Convection-allowing model (CAM) ensembles produce valuable information on convective mode and evolution that coarser, convection-parameterizing models cannot provide. This information aids forecasters in formulating severe weather outlooks at lead times of 1 to 2 days. At longer lead times, there is not an operational ensemble CAM capacity available in the United States, so forecasters typically rely on environmental information from global models and ensembles to assess severe weather potential. It is hypothesized that if CAM ensembles were available at longer lead times, then severe weather forecasts would be improved over the current environment-only approach. This hypothesis is evaluated by analyzing output from two global ensemble systems that provide output at 6-hourly increments out to 30 day lead times. One ensemble uses a uniform 13-km mesh over the globe, while the other uses the 13-km global grid with a 5-km convection-allowing nest over the contiguous United States (CONUS). Random forests trained to predict severe weather over the CONUS using the environmental information from the uniform mesh are compared to those trained using the environmental plus storm-attribute information from the global-nest ensemble configuration.

The model data was provided by GFDL and uses their System for High-resolution Prediction on Earth to Local Domains (SHiELD), which is based on the Finite-Volume Cubed-Sphere Dynamical Core (FV3). Daily forecasts out to Day 15 are generated in the cool season (October – May) from 2016 – 2020 and evaluated against observed severe weather. Results show that using the CAM nest ensemble and its simulated updraft helicity improves severe weather forecast skill through forecast Day 8. By Day 10, both the Global and the CAM nest ensembles show little skill at predicting severe weather.