A feedforward neural network (NN) was trained to produce gridded probabilistic convective hazard predictions over the contiguous United States. Input fields to the NN included 174 predictors, derived from 38 variables output by 497 convection-allowing model forecasts, with observed severe storm reports used for training and verification. These NN probability forecasts (NNPFs) were compared to surrogate-severe probability forecasts (SSPFs), generated by smoothing a field of surrogate reports derived with updraft helicity (UH). NNPFs and SSPFs were produced each forecast hour on an 80-km grid, with forecasts valid for the occurrence of any severe weather report within 40 or 120 km, and 2 h, of each 80-km grid box. NNPFs were superior to SSPFs, producing statistically significant improvements in forecast reliability and resolution. Additionally, NNPFs retained more large magnitude probabilities (> 50%) compared to SSPFs since NNPFs did not use spatial smoothing, improving forecast sharpness. NNPFs were most skillful relative to SSPFs when predicting hazards on larger-scales (e.g., 120 km vs. 40 km) and in situations where using UH was detrimental to forecast skill. These included model spin-up, nocturnal periods, and regions and environments where supercells were less common, such as the western and eastern United States and high-shear, low-CAPE regimes. NNPFs trained with fewer predictors were more skillful than SSPFs, but not as skillful as the full-predictor NNPFs, with predictor importance being a function of forecast lead-time. Placing NNPF skill in the context of existing baselines is a first-step towards integrating machine learning-based forecasts into the operational forecasting process.
NCAR is sponsored by the National Science Foundation.