Montgomery Flora - Mar 13

Name:     Montgomery Flora

Title:    Using machine learning to improve probabilistic forecasts of severe weather in the Warn-on-Forecast System

Location: NWC 5600

Date:     2020/03/13

Time:     3:00 PM

Series:   Convective Meteorology (Mesoscale Dynamics)

Abstract: A goal of the NOAA Warn-on-Forecast project is to provide rapidly-updating probabilistic guidance to human forecasters for short-term (e.g., 0-3 h) severe weather forecasts.  An ensemble of convection-allowing model forecasts can provide useful information on the likelihood of severe weather, but does not explicitly predict severe weather hazards (e.g., wind gusts, hail, and/or tornadoes). Machine learning algorithms can be applied to ensemble forecast output to produce probabilistic severe weather forecast guidance that improves upon that provided by the ensemble alone.

In this study, separate machine learning models are trained to predict the probability of 30-min-long ensemble forecast storm tracks (Flora et al. 2019) of producing a tornado, severe hail (> 1 in), and severe wind (>50 kts) using local storm reports as ground truth. Our dataset includes WoF system (WoFS) ensemble forecasts available every 5 minutes out to 3-h lead times from the 2017-2019 NOAA Hazardous Weather Testbed Spring Forecasting Experiments (86 dates). The training features include intra-storm variables, storm properties, and near-storm environment properties extracted from the forecast objects. We examine the performance of three machine learning algorithms: random forests, gradient-boosted trees, and logistic regression. All three models tend to discriminate well for all three hazards while producing fairly reliable predictions. Overall, the tree-based methods are significantly better than logistic regression. Additional details on the behavior and interpretability of tree-based methods will be presented.