Sam Varga

 Thesis Defense  

Creating Grid-Based Machine Learning Severe Weather Guidance for Watch-to-Warning Lead Times in the Warn-on-Forecast System

Sam Varga 

Thursday, July 18th, 2024 

NWC 5930 / 10:00 am 

If unable to attend in person join Google Meet:  

https://meet.google.com/hzj-xasw-onq 

Abstract: The Warn-on-Forecast System (WoFS) is a rapidly updating convection-allowing ensemble focused on providing numerical guidance at Watch-to-Warning lead times (0-6 hours). Previous studies (e.g. Flora et al., 2021; Clark and Loken, 2022) have incorporated machine learning (ML) to take advantage of the unique benefits of the WoFS and produce skillful guidance for severe weather hazards at lead times of 0-3 hours. This study evaluates the use of multiple ML algorithms to produce 2-6 hour severe weather guidance using data from the WoFS. This represents the first use of machine learning to produce WoFS-based guidance at these lead times and the first use of deep learning to produce severe weather guidance using WoFS data. 
 
Predictors are created using WoFS forecasts from the 2018-2023 Hazardous Weather Testbed Spring Forecasting Experiments. Data from forecast hours 2 through 6 are processed into predictors of multiple scales, incorporating both storm and environmental fields. We utilize three ML architectures: logistic regression, histogram-based gradient boosting trees, and U-nets. These models are trained to predict severe wind, severe hail, tornadoes, or any severe hazard during the 2-6 hour window. Target data comes from the NOAA Storm Events database. The four-hour ML guidance is compared to rigorous baselines consisting of optimized Neighborhood Maximum Ensemble Probabilities for each hazard. 
 
All ML methods evaluated outperform the NMEP baselines with tree-based methods achieving the highest performance. The largest improvement over the baseline occurs for severe wind guidance, followed by severe hail, tornado, and any severe guidance. Feature ablation shows that skill primarily comes from the intrastorm predictors and that the inclusion of multi-scale features exhibits little effect on skill. Despite the inclusion of additional features, the U-nets are unable to surpass the skill of the tree-based architectures. Similar to prior studies, this work shows the benefits of using the WoFS and ML to produce skillful guidance during the Watch-to-Warning period.