BoltCast: Deep Learning for Long Term Lightning Prediction

Lightning's threat to life, property, and infrastructure continues to increase in a changing climate. Many studies to date have focused on short term lightning prediction, and how to parameterize lightning flashes within traditional numerical weather prediction models. Few studies have investigated the correlation of convective parameters and the number of lightning flashes on a week time scale or longer, and none have used deep learning, or artificial intelligence (AI), to predict lightning on week time scales. This study aims to prototype a capability to use deep learning, to classify lightning probabilities out to 192 hours, or 8 days, using the Global Forecast System, or GFS model. Graupel, ice, rain, and snow mixing ratios, upward hydrostatic vertical velocities, convective available potential energy, composite reflectivity, and lifted index were used as model inputs over the conterminous United States (CONUS). GFS predictions were selected every 3-hours, from 00Z initializations to generate 8 different lightning probability products. All these input parameters generate a critical success index of .36 for the 00-hour specification. Results show that including the mixing ratios or the radar and instability data increased the critical success index by .1. Further metrics for different model architectures, and model inputs will be presented.