A Deep Learning Approach to Severe Weather Subseasonal Forecasting Over the United States

The deficiency in predictability at subseasonal timescales relative to that at conventional weather prediction timescales is significant. With respect to severe weather over the United States, previous work has shown promise for prediction at longer timescales through consideration of various teleconnections. For example, tropical modes of variability like the Madden Julian Oscillation have been shown to modulate severe weather frequency in the weeks following specific phases. Also tied to the occurrence of severe weather are synoptic features like the jet stream, deep upper-level troughs, and tropopause polar vortices. However, many of these processes and teleconnections have been exclusively investigated in the context of severe weather outbreaks over the United States with little consideration of interference between them.

Our work utilizes deep learning to improve severe weather prediction of tornadoes, hail, and wind on weekly to subseasonal timescales through examination of processes across multiple timescales and regions. We employ convolutional neural networks (CNNs) to identify periods of enhanced severe weather likelihood over different regions across the United States. The practically perfect severe weather hindcasts, Climate Data Record tropical outgoing longwave radiation anomalies, and ERA-5 synoptic and mesoscale data are used to train and test our model. Investigation of enhanced forecast skill with the use of explainable AI gives insight into the features and teleconnections strongly tied to longer-term predictability of severe weather hazards like tornadoes, hail, and wind.