Machine Learning-Based Convective Hazard Guidance Using Convection-Allowing Ensembles at Medium-Range Lead-Times

During Spring 2023 and 2024, convection-allowing ensemble (CAE) forecasts were produced in support of the NOAA Hazardous Weather Testbed (HWT) using the System for High-resolution prediction on Earth-to-Local Domains (C-SHiELD), a model based on the FV3 dynamical core. A novel aspect of these forecasts was their integration length well into the medium-range (out to Day 8), beyond the forecast lengths of operational CAEs, as well as their domains which spanned the entire globe, with convection-allowing grid spacing over the contiguous U.S. (CONUS). This demonstration was the first of its kind to assess forecast value from CAEs at these medium-range forecast lead-times.

Here we focus on evaluating probabilistic guidance for convective hazards (e.g., tornadoes and hail storms) at medium-range lead-times using the CAE output. Given the success of using machine learning (ML) techniques, such as neural networks (NNs), to post-process and generate convective hazard predictions at short lead-times (Day 1–2), we extended those techniques to generate forecasts at Days 3–8, using the C-SHiELD forecast dataset. Specifically, NNs were trained to predict the probability of severe weather reports across CONUS in 4-hr periods through Day 8 using a training dataset of C-SHiELD forecasts from Spring 2022. Then, NN predictions were made using the Spring 2023 and 2024 forecasts and evaluated with a mix of objective and subjective methods. We also generated convective hazard guidance with the updraft helicity (UH) diagnostic to serve as a non-ML forecast baseline.

The NN-based severe hazard predictions were skillful relative to climatology out to Day 8, with skill decreasing with lead-time. The NN forecasts were also skillful relative to the surrogate UH forecasts, although the difference between the ML and non-ML baseline decreased with lead-time, approaching zero by Day 6. Participants in the 2024 HWT rated the NN guidance slightly lower than ML-based medium-range forecasts generated with the GEFS. We will report on ongoing investigations exploring these differences, including the use of a much smaller training dataset in the current work and the procedure used to convert 4-hr probabilities into 24-hr probabilities, which were natively output by the GEFS ML systems. Finally, several example cases will be shown where the NN-based forecasts were markedly different from the underlying C-SHiELD forecasts at medium-range lead times, offering evidence that the NNs were capable of improving the spatial placement of severe hazard probabilities compared to the raw forecast output.