Evaluation of Real-Time Machine Learning Hail Forecasts from the NCAR Convection-Allowing Ensemble

The National Center for Atmospheric Research (NCAR) has been running an ensemble of WRF models at convection-allowing resolution daily since April 2015. By running in a fixed configuration for over a year, the NCAR ensemble has provided a large and diverse dataset for training and evaluating post-processing algorithms on convection-allowing models. This project used the NCAR Ensemble output to train random forest models to predict hail occurrence and hail size distributions for individual storms. The machine learning hail forecasts were compared with hail size forecasts from HAILCAST, estimates of hail size from the microphysics, and the intensity of storm-surrogate variables, such as updraft helicity.

The random forest models were trained using NCAR ensemble runs from May through July 2015. Observations of hail size were gathered from the NOAA NSSL Multi-Radar Multi-Sensor radar mosaic Maximum Expected Size of Hail (MESH) product. Individual hailstorm tracks were identified in each ensemble member and were matched with MESH tracks. Information about storm intensity and environmental conditions was extracted from within each storm track. Separate random forest models predicted whether or not hail would occur and the parameters of each storm’s hail size distribution. Hail forecasts were generated in real-time for May through September 2016 and were displayed on the NCAR ensemble website.

Results have shown that the machine learning hail forecasts can discriminate reliably between storms that produce hail and those that do not. Hail size distribution parameter forecasts showed little bias and good sharpness. Storm-surrogate ensemble probabilities derived from the machine learning models generated fewer false alarms and higher probabilities of detection than other hail forecasting methods, particularly at the 50 mm threshold. Case studies and regional trends are also investigated.