This study tests the utility of convolutional neural networks (CNN) as a postprocessing framework for improving the National Center for Environmental Prediction’s Global Forecast System’s (GFS) integrated vapor transport (IVT) forecast field in the Eastern Pacific and Western United States. IVT is the characteristic field of atmospheric river (AR) events, which provide over 65% of yearly precipitation at some U.S. west-coast locations. When compared to GFS, the method reduces full field root mean squared error (RMSE) at forecast leads from 3 hours to 7 days (9-17% reduction), while increasing correlation between observations and predictions (0.5-12% increase). An RMSE decomposition shows that random error is predominantly reduced. Systematic error is also reduced up to 5-day forecast lead, but accounts for a smaller portion of RMSE. This work demonstrates that CNNs have the potential to improve forecast skill out to 7 days for precipitation events affecting the western U.S. This dataset includes: the calculated IVT from the National Center for Environmental Prediction’s GFS forecasts; the IVT from the National Aeronautics and Space Administration’s Modern-Era Retrospective Analysis for Research and Applications version 2 (MERRA-2) reanalysis; the Machine Learning Models used to Postprocess the GFS IVT. All data sets are for the North American West Coast and the Eastern Pacific.
View this collection on the contributor's website.