Extended Range Machine-Learning Severe Weather Guidance Based on the Operational GEFS

Research by Hill et al. (2023) demonstrated promising results using a random forest (RF) machine-learning algorithm with input from the Global Ensemble Forecast System v12 (GEFSv12) to generate probabilistic severe weather forecasts out to days 4-8. In their work, the GEFS Reforecast Dataset (GEFS/R) was used to train and test an RF model, which was used to generate forecasts using operational GEFS forecasts as input. One limitation of Hill et al. (2023) was that, due to computational limitations, the GEFS/R forecasts only include 5 members. This work builds on Hill et al. (2023) by formulating a similar RF model using the operational GEFSv12 dataset for training and testing, instead of the reforecasts, to produce severe weather probabilities at 1-15 day lead times. At the time of this writing, GEFSv12 has been operational for more than three years, so it may be possible to use all 31 members in the operational GEFSv12 to get an improved result.

Before RFs trained from operational and reforecast GEFS data are compared, general performance characteristics are assessed from a control RF using ensemble mean predictors from operational GEFS forecasts. Then, for optimization and to assess performance sensitivities, feature engineering experiments are conducted to gain insight on the impact of ensemble member number, ensemble summary measures, and a multi-model strategy where predictors from both GEFS and the European Center for Medium Range Weather Forecast’s Integrated Forecast System are combined.

The feature engineering experiments found only modest degradation in forecast skill with decreasing numbers of members. The algorithm performed best using simple ensemble means as predictors rather than three percentiles (25, 50, and 75%) or individual members, and the multi-model RF resulted in modest forecast improvement. Using the optimal RF configuration, an RF trained on operational GEFS forecasts performed significantly better than an RF trained on reforecasts. Additionally, it was found that the RF described in Hill et al. (2023) performed slightly worse than the RF in this study trained on reforecasts. We speculate that superior performance in the RF trained from operational data could be related to the operational GEFS forecasts themselves having better performance than the reforecasts, and/or the storm report database better sampling severe weather in more recent times relative to the 2000-19 period.