969 A Machine Learning Approach to Severe Thunderstorm Downburst Prediction across Phoenix, Arizona

Tuesday, 14 January 2020
Hall B (Boston Convention and Exhibition Center)
Luke LeBel, Univ. at Albany, SUNY, Albany, NY; and P. M. Iniguez and J. Rogers

Severe thunderstorm wind events in the Phoenix, Arizona metro area are challenging to predict in advance due to their small spatial extent and short temporal duration. Moreover, the rapidly growing population and large spatial extent of the Phoenix metro area continues to increase the risk to life and property in association with severe wind events. It is therefore increasingly important to understand the signatures that precede the occurrence of damaging winds and improve their prediction.

This study attempts to address this problem by developing a predictive tool that can be used in support of severe weather operations at the NOAA/NWS Weather Forecast Office (WFO) in Phoenix, AZ. A dataset of radar and environmental attributes for 85 thunderstorm cases (30 severe thunderstorms and 55 sub-severe thunderstorms) was manually developed. A random forest (RF) algorithm was then trained on this dataset. Preliminary results have shown varying distributions in radar characteristics discriminating between severe and sub-severe thunderstorm cases, suggesting that a RF algorithm may be able to skillfully distinguish between these classes. Radar attributes such as echo top heights and the maximum vertically integrated liquid (VIL) had the strongest distinctions. Testing of the RF algorithm suggests operational implementation of a predictive model to successfully detect severe wind events may be possible, but will first require real-time tests and a more rigorous evaluation of the RF algorithm.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner