Machine Learning Probability Ensembles for Medium-Range Excessive Rainfall Prediction

Machine learning (ML) has provided a new avenue by which we can generate predictions of excessive rainfall with sufficient lead time to inform decision makers and planners of the threat of inclement weather. ML techniques are able to encode known long-standing relationships between environmental predictors and convective hazards from long historical records, and they have demonstrated tremendous value in predicting weather hazards at longer lead times. Further, continued effort by the meteorological community to explain ML models and their forecasts is building trust between developers and end users. As a result, their use in meteorological hazard forecasting is expanding, particularly into the medium range (e.g., 4-8 days) when forecasters are reliant on relatively coarse NWP models to create forecasts.

One such example is the ML system developed at Colorado State University, which is configured with Random Forests (RF) and has transitioned into operations at the Weather Prediction Center. The ML system forecasts excessive rainfall into the medium range by leveraging median input fields from the Global Ensemble Forecast System (GEFS), demonstrating predictive skill out to day 6. However, the value of an ensemble of RF forecasts for excessive rainfall has yet to be explored. It is hypothesized that a distribution of probabilistic forecasts could help to extend forecast skill beyond 8 days, in an analogous way to how ensemble frameworks have transformed traditional NWP from deterministic solutions. In this work, we extend the RF-based system to make forecasts of excessive rainfall with each individual GEFS ensemble member, thereby generating a distribution of probabilistic forecasts. Each probabilistic forecast maps to a categorical definition (e.g., marginal, slight, moderate, and high-risk categories), so the distribution of RF forecasts represents a distribution of categorical outlooks, which forecasters can use when uncertain about including a higher category in their human-generated and public-facing outlooks. Initial qualitative evaluations suggest moving away from deterministic RF-based forecasts towards ensemble forecasts yields increased awareness of excessive rainfall potential, with slight risk categories (i.e., 15-40% probabilities) being identified 7 days in advance of Tropical Storm Debby in South Carolina which didn’t exist in the operational ML system. In this presentation, we will highlight performance characteristics of the ensemble of RFs during specific high-impact rainfall events and detail aggregate performance over nearly four years of reforecasts.