Progress toward forecasting excessive rainfall with random forests based on a deterministic convection-allowing model

Machine learning applied to high-resolution numerical weather prediction forecasts is gaining popularity for prediction of high-impact weather events, but questions remain on how to design such systems. Building off prior work, we describe sensitivity experiments using a random forest (RF) system for excessive rainfall prediction using the deterministic High-Resolution Rapid Refresh (HRRR). We find that applying an RF trained on HRRRv3 to daily forecasts from HRRRv4 does not significantly degrade results. In terms of the spatial aggregation of predictor information for the RF, we find that forecasts are significantly improved when the system uses data from all nearby model gridpoints, rather than just sparse input points, particularly for events featuring small-scale precipitation maxima. Forecasts are also improved when predictors are averaged across multiple HRRR initializations rather than taken from a single HRRR initialization, as the RF is better able to learn relationships with the reduced uncertainty of the time-lagged ensemble mean predictors. The HRRR-based RF was evaluated during the Flash Flood and Intense Rainfall (FFaIR) Experiment over three consecutive years, with year-over-year improvements in the system. The HRRR-based RF exhibits an underprediction bias, partially attributable to using a more comprehensive dataset of excessive rainfall for verification than for RF training. Evaluation of RF skill during more than four years of real-time forecasts reveals that RFs based on deterministic convection allowing models continue to underperform relative to a Global Ensemble Forecast System based RF, and neither RF is as skillful as the operational excessive rainfall outlook for the day one period.