Applying Dense and Convolutional Neural Networks to the HRRR and RRFS to Generate Reliable, Skillful 1-48 Hour Predictions of Severe Weather Reports, Warnings, and Lightning

Machine learning (ML) techniques have been extensively used to post-process numerical weather prediction (NWP) output to generate reliable, skillful, predictions of severe weather hazards at lead-times from hours to days. Here, we applied a dense neural network (DNN) and a convolutional neural network (CNN) to generate hourly convective hazard predictions using output from three convection-allowing models (CAMs): HRRRv3, HRRRv4, and the experimental Rapid Refresh Forecast System (RRFS). We trained the DNNs and CNNs using HRRR forecasts from 2018–2020 and applied the trained models to forecasts from 2021–2022. The ML models predicted the probabilities of different report types (e.g., hail > 1”, wind gusts > 50 knots, and tornadoes), the occurrence of NWS severe thunderstorms and tornado warnings, and remotely-sensed lightning flashes. Training with NWS warnings provides a potentially useful alternative to storm reports, e.g., tornado warnings can be used as a proxy for the occurrence of low-level rotation and severe thunderstorm warnings can be used as an indicator of intense convection, without introducing storm report biases.

First, we validated the 2021–2022 HRRR-based DNN predictions to quantify the practical predictability of convective storms and their hazards, as a function of both initialization time and hour of the day. Overnight predictions were less skillful than daytime predictions, and afternoon HRRR initializations (e.g., 18 – 21 UTC) improved hazard forecast skill, relative to earlier initializations likely as a result of radar data assimilation. This overnight minimum in skill is likely related to model errors or predictability limits, more so than an issue with storm reports, given a similar diurnal cycle is observed in the lightning predictions. We also explored how predictions from the DNN and CNN differ in terms of their skill and reliability. While both systems produced probabilistic forecasts with large AUCs for all hazards, the CNN system tends to be less calibrated. Work is ongoing to improve forecast calibration for the CNN and to determine if an optimal combination of the two systems outperforms each individually. Given the eventual transition of the HRRR to the RRFS, we plan to apply the DNNs and CNNs to produce predictions of hazards during Spring 2023 using experimental RRFS output. We will explore the severe weather prediction capabilities of the RRFS using these ML-based forecasts.