Forecasting Tornadoes with Machine Learning

The destructive nature of tornadoes results in a significant number of fatalities and costs the U.S. billions of dollars each year. Even with advances in numerical weather prediction, forecasters are still only able to provide short term notifications to the public about the potential for tornadoes. Machine learning models are being trained in the prediction of extreme weather events, and this work applies these techniques to the prediction of weather conditions that are favorable for the development of tornadoes.

The National Oceanic and Atmospheric Administration's (NOAA) storm event database was used to generate a list of tornadoes from the years 2006 to the present. Using the space and time location of the events, atmospheric reanalysis data from the Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2) was used to generate a large training set of one-hour time averaged variables, such as wind speed, temperature, humidity, and other variables at multiple levels in the atmosphere at the time and location of the tornado. This data set was divided between 80% training and 20% validation, and the models were trained to a very high degree of accuracy. Using multiple models allows for sensitivity analysis for the input variables to the different models. In addition, this presentation will provide an analysis of the predictions of tornadoes using these trained models for the time periods between 2017 and the present and provide an analysis of the accuracy of the predictions. Moving forward, the use of high resolution forecast data will improve the accuracy of the predictions, and these methods could be used to predict other extreme weather events.