Heavy Precipitation and Flash Flood Forecasts Using Random Forests and Convection-Allowing Models

A random forest (RF) machine-learning model is currently implemented at the NOAA Weather Prediction Center to produce probabilistic forecasts of extreme rainfall in the Day 2 (36-60 hours) and 3 (60-84 hours) timeframes. These forecast products are based on NOAA’s Second-Generation Global Ensemble Forecast System Reforecasts (GEFS/R), 11-member coarse-resolution global model simulations, that provide predictors to train against observed events of flash flooding. Recent efforts have been focused on producing similar forecasts for the day-1 (12-36 hour) time period, but it is not yet clear whether using the GEFS/R (which has coarse resolution but a longer record) or convection-allowing models (CAMs, which can explicitly represent convective storms that produce heavy rainfall but have frequent changes in configuration) will yield the best results. To this end, deterministic (e.g., High Resolution Rapid Refresh (HRRR), National Severe Storm Laboratory WRF (NSSL-WRF)) and ensemble CAM varieties (e.g., HRRR-ensemble) are explored and compared to a GEFS/R based system to generate reliable Day-1 forecast guidance with RFs.

Sensitivity experiments are designed to methodically configure the RF for Day 1 forecasts using CAMs. The RF has a number of tunable parameters, including predictor selection (e.g., variable type, spatial and temporal selection), training period length, preprocessing (e.g., principal component analysis), and verifying-observation type (e.g., flash flood reports). Experiments are configured to systematically evaluate the parameters to create an optimal, accurate, and reliable probabilistic forecast. Of particular interest is the value of ensemble information to extend the training period length, which is typically small for CAMs given they frequently undergo model upgrades. The experimental results will be presented and discussed in the context of improving the RF algorithm for Day 1 probabilistic flash flooding guidance.