Noting that more than half of the power outages in the U.S. are estimated to have occurred from severe storms, an Electric Power Research Institute study explores the application of a novel machine learning method to reproduce historical tornado patterns and project future tornado activity.
The study, which was released in February, was done in collaboration with Kent State University.
“Tornadoes are far too small to be simulated directly by global climate models, and even the storms that produce them are not well-resolved by those models,” the study said. “But given their impact on power systems, it’s worth exploring alternative methods that can be used to estimate potential changes to tornado patterns in a warmer world."
“By leveraging the strengths of climate models in simulating large-scale atmospheric patterns, this research identifies large-scale patterns important for tornado formation and uses these patterns to predict future tornado activity, effectively side-stepping a major limitation of climate models in capturing smaller-scale tornado-specific variables,” the report said.
To develop tornado projections, the study applies a methodology called synoptic typing that includes principal component analysis and self-organizing maps to identify atmospheric patterns associated with tornado days.
Key Findings
The report details three key findings as follows:
- Machine learning can effectively reproduce tornado patters with limited inputs
- The capability of machine learning to project future conditions is generally limited by the climate models representation of large-scale circulation patterns.
- Projections using machine learning suggest an increase in tornado activity in the Midwest and a shift in tornado timing toward early spring and winter.
