Coupling Convection-Allowing Model Guidance with Machine Learning for Day-Ahead Extreme Precipitation Outlooks

The random forest (RF) algorithm is implemented to produce a day-ahead probabilistic forecast product across the contiguous United States (CONUS) for risk of locally extreme, flash-flood producing precipitation analogous to the Excessive Rainfall Outlook (ERO) issued by the Weather Prediction Center (WPC). Output from over six years of National Severe Storms Laboratory’s Weather Research and Forecasting model (NSSL-WRF) from between June 2009 and August 2016 are used as input to the RF. Separate forests are trained for each of eight hydrometeorologically distinct geographic regions and for three different time intervals: 1) a 24-hour forecast spanning 1200-1200 UTC, 2) a 6-hour forecast spanning 1800-0000 UTC, and 3) a 6-hour forecast spanning 0000-0600 UTC. An investigation of how to most effectively use high-resolution model output for predictor construction is conducted. Comparison with using a longer record of forecasts from NOAA’s Second Generational Global Ensemble Forecast System Reforecast Version 2 (GEFS/R) for RF training, as has been previously employed for Day 2-3 ERO guidance, is also made. Using both GEFS/R and NSSL-WRF guidance for predictors in a single model is also explored. Consistent with practice at WPC, models use incorporate several sources into their predictand and for verification, including average recurrence interval exceedances and flash flood local storm reports. Regional forecasts are blended to produce a consistent, CONUS-wide probability field each day. These probability fields are evaluated for skill and reliability over a withheld 2 year evaluation period spanning September 2016-August 2018. Preliminary versions of these models were evaluated in summer 2018 during the Flash Flood and Intense Rainfall Experiment (FFaIR) at WPC. Quantitative results from the 2-year verification as well as subjective evaluation and feedback of these models from FFaIR will be presented.