Skip to main content
U.S. flag

An official website of the United States government

Predicting Damaging Compound Freezing Rain-Gust Events Using Machine Learning

Presentation Date
Friday, December 16, 2022 at 2:45pm - Friday, December 16, 2022 at 6:15pm
Location
McCormick Place - Poster Hall, Hall - A
Authors

Author

Abstract

Based on our analysis of NOAA Storm reports, 533 significant ice storms were reported in CONUS during 2005-2021. Compound atmospheric hazards such as freezing rain events, coupled with high wind gusts (FZ-G), can cause largescale damage to civil and energy infrastructure and pose a threat to road and aviation safety. These effects are magnified when wind gusts occur shortly after significant ice accumulations build up. In fact, large ice loads make even weak and moderate gusts much more likely to cause widespread disruptions. Based on our analyses of data from over 700 ASOS stations these compound events occur throughout much of the USA and are particularly frequent in the Midwest and Northeast. One such event during December 2015 was associated with a strong mid-latitude cyclone which tracked from the Gulf of Mexico. During the initial phase, freezing rain and strong wind gusts caused travel disruptions in Texas while later, freezing rain led to the build-up of up to 0.5 inches of ice in the Indiana-Ohio-Michigan corridor which, in conjunction with strong wind gusts (up to 27 m/s), caused power outages for hundreds of thousands over several days.

While past work has focused on predicting the occurrence of strong and damaging wind gusts or freezing rain in isolation, there has been less work done on predicting compound freezing rain-gust events. Part of the challenge to forecasting FZ-G events lies in the complex factors which drive them. They are associated with very specific, and hard to predict, atmospheric conditions with small variations in thermal profiles, local scale stability and surface energy balance, as well as the timing and phase of precipitation, yielding very different threat levels. These complexities represent both a challenge to forecasting but also a problem that might be amendable to machine learning frameworks.

In this work we provide a comprehensive climatology of FZ-G events over CONUS using ASOS records and the NOAA Storm Reports for 2005-2021 and then apply artificial neural networks to predict FZ-G events at each station and then at an aggregated spatial scale using predictors from ERA5. Models are assessed for their balance of accuracy and complexity, with the ideal goal of producing a generalizable model-form and using it both for developing insights into the processes and for more accurate forecasts.

Category
Atmospheric Sciences
Funding Program Area(s)