Using Machine Learning to Improve Storm-Scale 1-h Probabilistic Forecasts of Severe Weather

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 probability of severe weather, but often contain large errors in storm intensity, timing, and location. Machine learning algorithms can be applied to ensemble forecast output and additionally incorporate diverse observational data (e.g., radar, satellite) to produce probabilistic forecast guidance that improves upon that provided by the ensemble alone.

In the first part of this study, hour-long storm tracks from the real-time Warn-on-Forecast System (WoFS) forecasts generated during the 2017-2018 NOAA Hazardous Weather Testbed Spring Forecasting Experiments were matched against observed rotation tracks, a proxy for low-level mesocyclones (a necessary precursor for tornadogenesis). Observed rotation tracks were derived from the NSSL Multi-Radar/Multi-Sensor (MRMS) low-level azimuthal shear analyses. Extracted information from the forecast storms include intra-storm variables, storm properties, and conditions of the forecast pre-storm environment. Multiple machine learning algorithms (e.g., random forests, XGBoost, logistic regression) including deep learning approaches (e.g., convolution neural network) were employed in this study. To demonstrate the skill of the various machine learning algorithms, forecasts were compared against the current accuracy and reliability of the WoFS probabilistic forecast of low-level rotation. All algorithms provided similar forecast accuracy and reliability as the WoFS, but were unable to outperform it. The algorithms are likely suffering from a lack of data and without additional, bias-correcting information cannot outperform the explicit CAM predictions of rotation for such short lead times. In future work, we plan to include observations (e.g., MRMS reflectivity) in the training data.

In the second part of this study, the hour-long forecast storm tracks will also be matched against local storm reports (e.g., wind gusts, hail, and tornados) in addition to observed low-level rotation to explicitly predict the likelihood of severe weather. A larger dataset spanning the 2016-2019 NOAA Hazardous Weather Testbed Spring Forecasting Experiments will be used. Preliminary results from this work will be presented alongside the low-level rotation prediction results described above.