Using Machine Learning to Predict Warm-Season Convection over Northeastern Colorado

Hourly probabilistic forecasts of categorical composite reflectivity (CREF) are produced for 1500-0900 UTC for each warm-season (April-September) day over northeast Colorado, far southeast Wyoming, and far southwest Nebraska. Forecasts from 0000 UTC initializations of the convection-allowing National Severe Storms Laboratory Weather Research and Forecasting model (NSSL-WRF) from all warm season days in 2013-2016 are used. Predictors come from thirteen atmospheric fields portrayed by the NSSL-WRF: composite reflectivity, convective available potential energy, convective inhibition, precipitable water, surface air temperature and dew point, ten meter zonal and meridional winds, six kilometer zonal and meridional winds, 500 hPa geopotential heights, planetary boundary layer height, and updraft helicity. Predictors include those displaced in both space and time relative to the forecast location to enable capability to correct for spatial displacement biases and convective initiation biases in the NSSL-WRF model. This procedure yields a large number of candidate predictors; these are then supplied to train neural networks (NNs) with varied configurations to make probabilistic CREF forecasts. The optimal NN configuration settings for this application will be presented, and commentary on the utility of deep NNs for this forecast problem will be made. Forecasts will be compared to those using simpler machine learning techniques, such as random forests (RFs), and straight-forward statistical techniques such as neighborhood-based techniques for creating forecast probabilities. Preliminary results suggest that these methods, with proper tuning, can produce significantly improved forecast skill compared to forecasts produced using traditional means.