Evaluating Probabilistic Predictions of Lightning Generated with a Neural Network and HRRR Forecasts

Extending previous work with neural networks and local storm reports, we have applied a similar machine learning (ML) technique to lightning observations. Using a year of 00 UTC initialized HRRRv4 forecasts, we trained several neural networks to predict lightning in 1-4 hour windows and 20-40 km spatial windows out to 48 hours given standard output like CAPE, updraft helicity, and simulated reflectivity. The models were trained and evaluated with observed cloud-to-ground and in-cloud flashes both from surface-based Earth Networks and GOES Lightning Mapper satellite data. Different flash thresholds were used to identify weak vs. strong convection. Overall, the ML lightning forecasts exhibited higher skill than the severe weather report forecasts at all time and space scales (1, 2, 4 hr window and 20, 40 km scales). This may be due to lightning's greater prevalence or predictability, or because the lightning observations suffer less from timing and location biases. Overall, the lightning predictions serve as a potentially useful guidance tool for users sensitive to the occurrence of thunderstorms, and are being generated in real-time for dissemination and evaluation. Future work will explore the differences in predictor importance between severe weather and lightning models and cases which exhibit either lightning or severe weather, but not both. The ML forecasts will also be compared to non-ML surrogate forecasts using lightning diagnostics, Storm Prediction Center (SPC) calibrated lightning products, and SPC operational forecasts of thunderstorms.