Next-day probabilistic severe weather predictions using convection-allowing models and convolutional neural networks
Point-based machine learning models, such as decision trees and multilayer perceptrons, have long been applied to predict severe weather hazards, such as tornadoes, large hail, and intense convective wind gusts. However, few works have post-processed convection-allowing model outputs with methods that can identify features and spatial patterns within two-dimensional forecast output. In this research, a Convolutional Neural Network (CNN) is developed to predict severe weather hazards at 1 – 36 hr lead times. The CNN is trained with convective-scale and environmental variables derived from the 3-km deterministic High-Resolution Rapid Refresh (HRRR). In addition to employing a CNN, this work differs from prior work by using the native 3-km horizontal grid spacing forecast output for training rather than upscaling to a coarser mesh, enabling the CNN to learn storm-scale information potentially related to convective hazard likelihood, such as convective mode. The CNN is designed to predict the probability of the occurrence of any severe weather report within 40-km and 2-hr of a point in the Continental United States using National Weather Service storm reports. The performance of the CNN-based predictions is evaluated via a withheld set of HRRR forecasts, as well as against UH-based surrogate-severe guidance and other machine learning models trained to produce probabilistic predictions of severe weather hazards. Objective verification metrics are used to assess the forecast resolution and reliability. Feature importance, sensitivity tests, and process-based analysis are conducted to explain the behavior of the CNNs in severe weather prediction problems. This work demonstrates that CNNs can potentially be used as post-processing tools for high-resolution convection-allowing models and produce robust predictions of weather hazards using storm-scale forecast information.