Using Machine Learning to Predict Storm Longevity in Real Time

Real-time prediction of a storm's duration is a critical task for forecasters. These predictions guide forecasters when they issue warnings and can inform them about the potential severity of a storm. An automated prediction system may also be able to help reduce forecaster cognitive load during high workload events.

In this work, we present a machine learning approach for automated real-time prediction of storm longevity. This system was tested in the National Oceanic and Atmospheric Administration (NOAA) Hazardous Weather Testbed (HWT) in the summer of 2016 and 2017, and specifically within the Probabilistic Hazard Information (PHI) warning software (Karstens et al. 2015).

Labels for the training data were provided by the best-track method from the Warning Decision Support Integrated Information (WDSII), and storm characteristics were drawn from data provided by the NOAA / Cooperative Institute for Meteorological Satellite Studies (CIMSS) ProbSevere model (Cintineo et al, 2014). We also experimented with radar data obtained from the Multi-Year Reanalysis of Remotely Sensed Storms (MYRORSS) and Near Storm Environment (NSE) datasets created by interpolating soundings from RUC/RAP archives to the storm location using SharpPy (Blumberg et al, 2017).

We demonstrate that the predictions are highly accurate using multiple machine-learning methods and discuss the preferred predictive biases of the human forecasters compared to the automated methods.