Developing an Automated System to Predict Tornadoes in Simulated Nonclassical Convective Storms

It seems inevitable that machine learning will become increasingly important in the detection and short-term prediction of severe thunderstorms. We explore the utility of using machine learning to predict tornadoes in an ensemble of 128 simulations performed using Cloud Model 1. The simulations are configured to produce storms that are “non-classical” in terms of their radar appearance, as opposed to the more classical signatures associated with supercell storms. Nonclassical storms are particularly challenging for forecasters trying to issue accurate warnings. The ensemble of simulations is produced using a variety of initiation methods and soundings, and results in a diverse set of storm modes that includes multicell storms and squall lines. Over 600 tornado-like vortices develop within the ensemble of simulations.

Throughout each simulation, storm objects are identified and tracked to collect information about the characteristics of each storm including features such as its updraft, cold pool, and near storm environment. The collected data are then used to produce multiple machine-learning models, each trained on forecasting tornadoes at different lead times. The combination of the output from multiple machine learning models is used to create a tornado forecasting system. The system is capable of providing the probability of a particular storm producing a tornado at 2-minute intervals, and making decisions on whether or not to issue tornado warnings. We will explore the feasibility of automated tornado prediction in nonclassical convection and present a framework for implementing a fully automated tornado forecasting system.