Machine Learning-Derived Severe Weather Probabilities from a Warn-on-Forecast System

Severe weather probabilities are derived from the Warn-on-Forecast System (WoFS) run by NOAA's National Severe Storms Laboratory (NSSL) during spring 2018 using the random forest (RF) machine learning algorithm. Recent work has shown this method generates skillful and reliable forecasts when applied to convection-allowing model ensembles for the "Day 1" time range (i.e., 12-36-h lead times), but it has been tested in only one other study for lead times relevant to WoFS (e.g., 0-6 h). Thus, in this paper, various sets of WoFS predictors, which include both environment and storm-based fields, are input into a RF algorithm and trained using the occurrence of severe weather reports within 39 km of a point to produce severe weather probabilities at 0-3-h lead times. We analyze the skill and reliability of these forecasts, sensitivity to different sets of predictors, and avenues for further improvements. The RF algorithm produced very skillful and reliable severe weather probabilities and significantly outperformed baseline probabilities calculated by finding the best performing updraft helicity (UH) threshold and smoothing parameter. Experiments where different sets of predictors were used to derive RF probabilities revealed 1) storm attribute fields contributed significantly more skill than environmental fields, 2) 2-5 km AGL UH and maximum updraft speed were the best performing storm attribute fields, 3) the most skillful ensemble summary metric was a smoothed mean, and 4) the most skillful forecasts were obtained when smoothed UH from individual ensemble members were used as predictors.