Postprocessing HREFv2 Heavy Rainfall Forecasts Using Machine Learning

The High Resolution Ensemble Forecast System Version 2 (HREFv2) became the first operational convection-allowing ensemble in November 2017. While it has generally provided skillful and valuable forecast guidance for fields related to convection, its raw forecast probabilities of heavy rainfall tend to be over-confident. Moreover, like many convection-allowing ensembles, its members suffer from biases in the placement and magnitude of forecast precipitation. One way to correct for these biases is through the application of machine learning (ML), which has shown promise as a post-processing technique in the past. However, the value of ML relative to more basic post-processing methods remains largely untested.

In this study, a random forest (RF) algorithm is used to produce probabilistic 12-36 hour precipitation forecasts for the contiguous United States using input data from the HREFv2. Predictor variables include member forecasts and ensemble statistics of a variety of fields, including: temperature and dewpoint temperature at multiple vertical levels, simulated 1 km above-ground reflectivity, surface-based CAPE and CIN, precipitable water, maximum hourly wind components, and forecast 24 hour precipitation. National Centers for Environmental Prediction (NCEP) Stage IV data are used as the observational dataset.

Results using data from late April 2017 to late March 2018 suggest that probabilistic 1-inch RF forecasts have excellent reliability and discrimination ability, especially compared to raw ensemble probabilities. However, RF forecasts are only marginally superior to spatially-smoothed raw ensemble probabilities. These findings demonstrate the importance of comparing ML results to meaningful baseline forecasts and suggest a need for large high-resolution ensemble datasets.