Generating Ensemble-Derived Next-Day Probabilistic Severe Weather Forecasts with Machine Learning

One challenge of severe weather forecasting is that numerical weather prediction guidance cannot explicitly simulate severe weather occurrence. Instead, proxies, such as simulated 2-5 km updraft helicity (UH), are typically used to infer if, when, and where severe weather will occur. A disadvantage of this approach is that the proxies must be calibrated, and the calibrations may be sensitive to geographic region and time of year. Additionally, proxies generally use only a small subset of available ensemble data, potentially ignoring valuable information. Machine learning (ML) can help solve these problems by directly relating a variety of ensemble forecast data to observed severe weather occurrence.

One ML method that has shown success in the past for post-processing ensemble-based precipitation forecasts is the random forest (RF). This presentation shows that, using a RF approach with data from 2015-2017, skillful probabilistic next-day severe weather forecasts can be obtained from the Storm Prediction Center’s (SPC’s) 7-member Storm-Scale Ensemble of Opportunity (SSEO) and the 10-member National Center for Atmospheric Research Ensemble Prediction System (NCAR EPS). The RF-based severe weather probabilities compare favorably with UH-based probabilities as well as the SPC’s day 1 convective outlooks, especially during the fall, when UH-based proxies are less skillful and severe weather forecasting skill reaches an annual minimum. It is envisioned that real-time RF-based probabilistic severe weather forecasts could help SPC forecasters bolster their day-1 convective outlooks by improving their access to information contained within high-resolution ensembles.