Kendall Junker - November 5

Convective Meteorology (Mesoscale Dynamics) Seminar

0-3 Hour Tornado Prediction Using the Warn-on Forecast System and Machine Learning

Kendall Junker

Friday, November 5th

3:00pm

Join Google Meet:

https://meet.google.com/iru-ggiv-afj

            This project will focus on creating key components of the research needed to generate operationally-useful tornado forecasts from the Warn-on Forecast System (WoFS) with machine learning techniques. Eventually, the algorithm will be evaluated in real-time at the Spring 2022 Hazardous Weather Testbed Spring Forecasting Experiment, then refined based on forecaster feedback.

            This research builds off prior machine learning research by Ryan Lagerquist at the University of Oklahoma. Lagerquist created a Convolutional Neural Network (CNN) to generate probabilistic tornado forecasts from real-time radar data. This research seeks to expand on Lagerquist’s work by applying similar techniques to forecasted radar data. By the end of the semester, a working machine learning model based upon Lagerquist’s codebase will be built and trained using at least 1000 GridRad storm events and will subsequently be tested using WoFS-derived radar data. GridRad storm data from 2011-2018 will be cut into folds for training and testing to improve the robustness of the model. For each fold, a year of data will be used as the testing data, and remaining data will be trained on. The ratio of training to testing data will begin at about 80% to 20%, but may be adjusted as necessary. To prevent synoptic correlation, there will be at least a two week buffer between date ranges of the tested and trained data.  Model effectiveness will be assessed using metrics such as binary cross-entropy, loss functions, the critical success index, and accuracy. The algorithm will be validated using a GridRad data set independent of the data utilized in the training and testing phase. The results will be used to optimize model design, look at change in forecast skill over time (i.e. the difference between a 0-30 minute forecast and a 150-180 minute forecast), and provide the outputs needed for NCAR’s visualization and XAI efforts.