Predictions of Severe Weather with Random Forests and the Global Ensemble Forecast System

Artificial Intelligence and Machine Learning techniques in meteorology have proliferated in recent years. Of particular interest are meteorological hazards -- tornadoes, large hail, and damaging winds -- that occur on spatial and temporal scales that are not well represented in numerical weather prediction (NWP) model output. The predictability limit for these hazards is short, so reliable probabilistic forecasts are needed rather than deterministic predictions that will inevitably have large errors. To address these challenges, over the past several years we have developed a suite of probabilistic forecast systems, referred to as Colorado State University-Machine Learning Probabilities (CSU-MLP), that use the Global Ensemble Forecast System (GEFS) Reforecast datasets, historical observations of severe weather, and machine learning algorithms to generate skillful, reliable guidance that operational forecasters can use as a "first guess" when generating outlooks. The CSU-MLP utilizes Random Forests (RFs) to generate probabilistic severe weather forecasts out to 8 days by training the RFs to learn how local environments relate to severe weather events. Nearly a decade of daily forecast initializations from the GEFS reforecast dataset are used to train the RFs and over two years of real-time forecasts are used to quantify forecast skill. This presentation will provide background on the CSU-MLP system, highlight the skill of the system in depicting severe weather events, and touch on the value added to the operational forecast process at U.S. national forecast centers, including feedback we have received from our operational forecast partners.