Characterizing and Forecasting Severe Convective Storms using Deep Learning

Abstract

This dissertation investigates the dynamics and predictability of tornadic supercell environments, leveraging advanced data analysis and machine learning techniques. Initial research established a foundation by examining the spatial and temporal variability of tornado-favorable parameters using reanalysis data. This work revealed substantial regional and temporal variations in these parameters, highlighting the limitations of using universal environmental thresholds for tornado prediction. Further investigation quantified the statistically significant differences between tornadic supercell and baseline environments, demonstrating that tornado-favorable conditions can exist well before storm formation, particularly outside of peak tornado seasons. This emphasizes the challenge of distinguishing between environments that produce tornadoes and those that do not. Building upon these foundational analyses, this research explored two interconnected avenues: the impact of climate change on severe convective storms (SCSs), and the application of deep learning (DL) for forecasting tornadic supercell environments. This research evaluated the performance of the deep learning model Pangu-Weather in forecasting tornadic environments one day in advance. Pangu-Weather's skill in predicting convective available potential energy (CAPE), 0-6 km shear, and 0-3 km storm-relative helicity was assessed and compared to the operational Global Forecast System (GFS). Results indicate that Pangu-Weather generally outperforms the GFS in predicting wind shear and helicity at the time and location of tornado reports, but tends to underpredict CAPE in the hours leading up to the event. To further enhance operational severe weather prediction, a novel neural network post-processing framework using decoder-only transformer was developed. This framework integrates forecasts from multiple models, including both numerical and AI-based models, and has the potential of accommodating various lead times from different models, improving the accuracy of SCS predictions. To investigate future SCS activity, a novel approach was developed using hierarchical clustering to identify potentially convective atmospheric profiles in both historical and future climate simulations. These profiles were then analyzed to understand the spatial and seasonal variations in potential SCS activity, providing a more nuanced analysis compared to previous studies relying solely on composite parameters. By combining data-driven analyses with advanced modeling techniques, this research provides valuable insights into the complex nature of tornadic supercell environments and their predictability in both present and future climates. These findings contribute to a deeper understanding of tornado dynamics and pave the way for improved forecasting capabilities, ultimately enhancing public safety and resilience to severe weather.