1A.4 Advances in Using Random Forests to Forecast Heavy Precipitation and Flash Floods

Monday, 8 January 2018: 9:30 AM
Room 17A (ACC) (Austin, Texas)
Gregory R. Herman, Colorado State Univ., Fort Collins, CO; and R. S. Schumacher

The random forest (RF) algorithm is implemented to develop skillful, calibrated contiguous United States (CONUS)-wide probabilistic forecasts of locally extreme precipitation, as quantified by a combination of average recurrence interval (ARI) exceedances, flash flood guidance (FFG) exceedances, and flash flood reports. Forecasts are made for two different 24 hour periods representing lead times of 36-60 hours and 60-84 hours. CONUS is partitioned into eight regions which exhibit similar hydrometeorological properties. Within each of these regions, forecasts are produced for each forecast point on a coarse (~0.5°) grid, each day in an 11-year historical period of record spanning 2003-2013, and for each of the two forecast intervals. Predictor data used to generate forecast probabilities come from an assortment of simulated atmospheric fields taken from a record of NOAA’s Second Generation Global Ensemble Forecast System Reforecast (GEFS/R) 11-member ensemble system, including not only the quantitative precipitation forecast (QPF) output from the model, but also variables that characterize the meteorological regime, including winds, moisture, and instability. For each field used, model forecast data is taken relative to each forecast point in space, and for each output time step over the given 24-hour forecast interval. This produces a large number of candidate GEFS/R predictors; to yield more tractable analysis and alleviate concerns of overfitting, this rearranged record of historical model data is pre-processed with principal component analysis (PCA), which is then supplied to an RF algorithm to produce locally extreme precipitation probabilities within a GEFS/R grid box. Recent innovations in the model development and real-time forecast procedures will be described and the results of applying this methodology presented. Overall, it is found that the RF-based forecasts add significant skill over exceedance forecasts produced from the raw GEFS/R ensemble QPFs across all regions of CONUS---more than is produced from 24 hours decrement in forecast lead time---and are quite reliable, particularly at small and moderate probabilities.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner