Advancing Probabilistic Forecasts of Locally Extreme Precipitation through Machine Learning

A statistical method is developed to produce skillful, calibrated contiguous United States (CONUS) wide probabilistic forecasts for extreme precipitation, framed in the context of average recurrence interval (ARI) exceedances, based off of a suite of different NWP models of varying and from varying sources. Specifically, forecasts are produced for 6 and 24 hour precipitation accumulations for event rarities ranging from 1-year to 100-year ARI exceedances.  A two step procedure is employed to produce these forecasts.  First, an assortment of surface and 3-D atmospheric fields are taken from models with very long historical model data records, such as NOAA’s Second Generation Global Ensemble Forecast System Reforecast (GEFS/R), to produce ARI exceedance probabilities on a coarse (~0.5°) CONUS-wide grid.  These large areal exceedance probabilities are then combined with quantitative precipitation forecasts (QPFs) from several convection allowing models (CAMs) such as the National Severe Storms Laboratory WRF (NSSL-WRF), High Resolution Rapid Refresh (HRRR), and 4km North American Mesoscale Nest (NAM-NEST) to produce probabilistic exceedance forecasts on a high resolution (~4.75 km) grid.  A variety of machine learning and other statistical techniques- including principal component analysis (PCA), random forests (RFs), support vector machines (SVMs), and gradient boosting- are employed to train these NWP model data and generate well calibrated forecasts.  Numerous developmental details, such as how statistical characteristics of extreme precipitation varies regionally, how statistical extreme precipitation analysis varies with season, and the choice of atmospheric variables to include in model construction, are explored.  Overall assessments of statistical model skill and reliability will also be presented.