Correcting, Improving, and Verifying Automated Guidance in a New Warning Paradigm
The prototype Probabilistic Hazards Information (PHI) system allows National Weather Service (NWS) forecasters to experimentally issue dynamically evolving severe weather warning and advisory products that provide hypothetical end users with specific probabilities that a given location will experience severe weather over a predicted time period. When issuing these products, forecasters are provided with an automated, first-guess storm identification object by the National Oceanic and Atmospheric Administration (NOAA) / Cooperative Institute for Meteorological Satellite Studies (CIMSS) ProbSevere model, which is intended to support the probabilistic warning issuance process. However, empirical results from experiments held at the NOAA Hazardous Weather Testbed (HWT) suggest that forecasters have a general distrust of the automated guidance, leading to frequent adjustments to the automated information.
Additionally, feedback from several years of experimentation suggest that forecasters have limited experience with how storm-scale severe weather probabilities tend to evolve in different convective situations.
To help address some of these concerns, the first part of this study provides a detailed analysis of the maximum attainable predictability of the automated ProbSevere guidance during the spring season of 2015, and offers a comparison of the verification statistics from automation to those of the corresponding storm-based warnings issued by the NWS during the same time period. To facilitate this process, a new algorithm is developed to correct improper track breakages that occur in the ProbSevere model, which could negatively affect usability, verification, and predictability metrics. The second part of this study addresses storm-scale severe weather probability trends by developing a machine learning model to predict the evolution of a storm’s likelihood of producing severe weather. This model uses the ensemble average of six machine learning members trained on variables obtained from the initial ProbSevere model, environmental parameters, and a storm’s history, to predict future ProbSevere probabilities over a predicted duration of a storm. The model was implemented and tested during the 2017 Spring Experiment held at the NOAA HWT, and an analysis of the model’s performance during that experiment is provided.