Improving Precipitation Forecasts with Convolutional Neural Networks

Abstract

Abstract A machine learning method based on spatial convolution to capture complex spatial precipitation patterns is proposed to identify and reduce biases affecting predictions of a dynamical model. The method is based on a combination of a classification and dual regression model approach using modified U-Net convolutional neural networks (CNN) to post-process daily accumulated precipitation over the US west coast. In this study, we leverage 34 years of high resolution deterministic Western Weather Research and Forecasting (West-WRF) precipitation reforecasts as training data for the U-Net CNN. The data is split such that the test set contains 4 water years of data that encompass characteristic west coast precipitation regimes: El Niño, La Niña, and dry and wet El Niño/Southern Oscillation (ENSO-neutral) water years. On the unseen 4-year data set, the trained CNN yields a 12.9-15.9% reduction in root mean square error (RMSE) and 2.7-3.4% improvement in Pearson correlation (PC) over West-WRF for lead times of 1-4 days. Compared to an adapted model output statistics correction, the CNN reduces RMSE by 7.4-8.9% and improves PC by 3.3-4.2% across all events. Effectively, the CNN adds more than a day of predictive skill when compared to West-WRF. The CNN outperforms the other methods also for the prediction of extreme events, which we define as the top 10% of events with the greatest average daily accumulated precipitation. The improvement over West-WRF’s RMSE/PC for these events is 19.8-21.0%/4.9-5.5% and MOS’s RMSE/PC is 8.8-9.7%/4.2-4.7%. Hence, the proposed U-Net CNN shows significantly improved forecast skill over existing methods, highlighting a promising path forward for improving precipitation forecasts.