Using Machine Learning to Improve Storm-Scale 1-h Probabilistic Low-Level Rotation Forecasts

One focus of the NOAA Warn-on-Forecast (WoF) project is to provide rapid-update probabilistic guidance to human forecasters for short-term (e.g., 0-1 h) tornado forecasts. Tornadoes are unresolvable with current operational models, but convection-allowing ensembles (CAEs) such as the 3-km NSSL Experimental WoF System for Ensembles (NEWS-e) provide forecasts of low-level rotation that may potentially discriminate between tornadic and non-tornadic storms. However, CAE forecasts often contain large errors in storm intensity, timing, and location. Machine learning methods can leverage ensemble uncertainty, incorporate several model variables, and mitigate forecast bias to produce calibrated probabilistic forecasts.

This study utilizes real-time NEWS-e forecasts generated during the 2016 -2018 NOAA Hazardous Weather Testbed Spring Forecasting Experiments and low-level azimuthal shear analyses from the NSSL Multi-Radar/Multi-Sensor (MRMS) product suite as input data for training various machine learning algorithms to produce calibrated 1-h probabilistic forecasts of low-level rotation. Three machine learning algorithms were used: random forests, gradient-boosted trees, and logistic regression. Given that strong low-level rotation occurs infrequently, we are oversampling the grid points within the rotational tracks to produce a more balanced dataset, which has been shown to improve machine-learning performance. In preliminary experiments, machine learning improves reliability and skill by bias-correcting the probability magnitudes, but without improving location prediction. Reducing phase errors is therefore the focus of ongoing work.