Convective Meteorology (Mesoscale Dynamics)

Machine-intelligent Prediction of Damaging Straight-line Storm Winds

Ryan Lagerquist

School of Meteorology

08 April 2016, 3:00 PM

National Weather Center, Room 5600
120 David L. Boren Blvd.
University of Oklahoma
Norman, OK

In the United States alone, thunderstorms cause over 100 deaths and 10 billion USD of insured damages per year. Many of these losses are caused by straight-line (non-tornadic) winds. Storm-scale winds are non-linearly related to many meteorological variables and are not adequately predicted by numerical models, which has motivated the use of machine learning (ML) for this problem.

We have developed a suite of ML models to predict the occurrence of damaging straight-line winds (> 50 kt), on a storm-by-storm basis, at lead times up to 60 minutes. These models use three types of input data: archived radar grids from the Multi-year Reanalysis for Remotely Sensed Storms (MYRORSS), proxy soundings from the Rapid Refresh (RAP) model, and surface wind observations from both weather stations and Storm Prediction Center (SPC) reports.

The first processing step is identifying and tracking storm cells from the MYRORSS data. Then wind observations are linked to nearby storm cells. Then four types of features are calculated for each storm cell: statistics on radar fields inside the storm cell; properties of the storm-cell shape; basic storm information (e.g., velocity and area); and sounding parameters from the RAP data. ML models (mainly random forests and gradient-boosted regression trees) are trained on these features to predict damaging winds. Finally, two methods (J-measure ranking and sequential forward selection) are used to rank the most important features, which helps provide physical insight into how the ML models are working.

Convective Meteorology (Mesoscale Dynamics) Seminar Series website