Predicting Storm Prediction Center Watch Likelihood Using Machine Learning

The National Weather Service’s (NWS) Storm Prediction Center (SPC) is responsible for issuing severe weather forecast products, including: Convective Outlooks, Mesoscale Discussions, and Severe Thunderstorm/Tornado Watches, which highlight locations where severe thunderstorms will be possible over the next several hours to days. Forecasters at the SPC issue Severe Thunderstorm and Tornado Watches once atmospheric conditions appear favorable for the development of severe weather, with the goal to provide one to two hours of lead time prior to the development of the first severe storm. Furthermore, Mesoscale Discussions are typically produced one to three hours prior to the issuance of a watch in order to alert NWS customers that organized severe thunderstorms and/or tornadoes may be possible in the near future. As such, SPC forecasters must start planning where, when, and for how long to issue watch products several hours before the first severe storm ever develops. This study serves as a preliminary investigation into using machine learning techniques to predict where a severe weather watch may be issued on a given convective day. In particular, tree-based learning methods will be trained on the SPC’s daily Convective Outlooks and individual hazard (tornado, wind, and hail) forecast probabilities to predict the likelihood that a watch will be needed at a given location during the convective day. Ultimately, the goal of this study and subsequent research is to produce an automated forecast guidance framework to help SPC forecasters better strategize where and when to issue Mesoscale Discussions and Severe Thunderstorm and Tornado Watches before hazardous weather develops.