Using machine learning to improve severe thunderstorm wind reports and diagnose environmental influences on measured severe gusts

Thumbnail Image
Date
2023-08
Authors
Tirone, Elizabeth
Major Professor
Advisor
Gallus, William A
Dutta, Somak
Newman, Jennifer
Gonzalez, Alex
Gutowski, William
Committee Member
Journal Title
Journal ISSN
Volume Title
Publisher
Authors
Research Projects
Organizational Units
Journal Issue
Is Version Of
Versions
Series
Department
Meteorology program
Abstract
The research encompassing the work of this dissertation looks to improve the usability of severe thunderstorm wind reports. There are many known limitations of the thunderstorm wind reports in the National Center for Environmental Information’s Storm Events Database. These limitations can have major ramifications on research methods that train or calibrate tools using these reports. The first study implemented a machine learning based tool to predict a probability that a severe thunderstorm wind report was caused by severe intensity winds (≥ 50 kts). Rigorous testing was done to improve the data and machine learning methods. This tool was shown at the Hazardous Weather Testbed Spring Forecast Experiment in 2020, 2021, and 2022 to gain feedback on its performance and implore additional techniques to improve performance. The output of this tool was used in the verification of Storm Prediction Center wind forecasts as well as in the creation of practically perfect hindcasts to allow for a proof of concept for potential uses. Following trends in machine learning to improve the explainability and interpretability of models, the second study aimed to improve the understanding of the tool from the first study. Using predictions from the best performing machine learning model, measured reports from 2018-2021 were evaluated based on their wind speed and assigned probability. Reports with a high wind speed but low probability and reports with a low wind speed but high probability were compared to reports with expected probabilities, so reports with a high (low) speed and high (low) probability. The findings from this study help improve end user confidence on the tool’s performance, and highlight difficulties with forecasting severe wind.
Comments
Description
Keywords
Citation
DOI
Source
Subject Categories
Copyright