Evaluation of Real-Time, Medium-Range, Convection-Allowing Ensemble Forecasts Produced for NOAA’s 2024 Hazardous Weather Testbed Spring Forecasting Experiment

As contributions to NOAA’s 2024 Hazardous Weather Testbed (HWT) Spring Forecasting Experiment (SFE), two sets of real-time medium-range (3–8-day) ensemble forecasts were produced with variable-resolution global models that had ~3-km horizontal cell spacing over the conterminous United States. One set of ~3-km ensemble forecasts had 10 members, produced 204-h (8.5-day) forecasts, and used GFDL’s “C-SHiELD”, a prediction model based on the Finite-Volume Cubed-Sphere Dynamical Core (FV3). The second set of ~3-km ensemble forecasts had 5 members, produced 132-h (5.5-day) forecasts, and used a variable-resolution configuration of NSF NCAR’s Model for Prediction Across Scales (MPAS). Both ensembles were initialized daily at 0000 UTC from Global Ensemble Forecast System (GEFS) initial conditions over 42 days encompassing the 2024 SFE. These forecasts, and similar demonstrations during the 2023 SFE, represented the first time convection-allowing ensemble guidance at medium-range forecast periods had ever been produced in real-time with variable-resolution global models. In addition, during the 2024 SFE, real-time medium-range convective hazards guidance was provided by a machine-learning (ML) system that was trained on C-SHiELD forecasts and then applied to the real-time C-SHiELD forecasts.

The forecasts were skillful and appeared to have value. For example, 2024 SFE participants found that both medium-range ensembles often successfully highlighted areas where severe weather occurred 3–7 days in advance. Additionally, accumulated precipitation forecasts from both ensembles were objectively more skillful than corresponding forecasts from NCEP’s much coarser operational GEFS through 3–4 days. The ML-based hazards guidance usually improved upon the raw C-SHiELD guidance.

This presentation will discuss these findings as well as lessons learned across both the 2023 and 2024 real-time demonstrations.