Improving Accumulated Precipitation Forecasts with Convolutional Neural Networks

Abstract:

Systematic biases in dynamical model forecasts mask valuable information about precipitation patterns in extreme weather events. Machine learning based post-processing methods can be used to provide more accurate and reliable predictions. In this study, we leverage 34 years of precipitation forecasts from a newly developed deterministic forecast model, West-WRF (WWRF). We use a U-Net architecture convolutional neural network to identify and reduce systematic spatial errors in the 24-hour accumulated precipitation over the North American west coast. The Parameter-elevation Relationships on Independent Slopes Model (PRISM) climate dataset is used as the ground truth and for forecast verification. The output of the U-Net CNN training and testing phases show a consistent 15% improvement in root mean squared error (RMSE) and a 10% improvement in mean absolute error (MAE) compared to the raw WWRF forecast. For severe precipitation events, we observe notable improvements in error metrics such as RMSE, BIAS and correlation. Hence, the proposed U-Net CNN shows improved forecast skill for precipitation events affecting the western United States. This study motivates research into a combined CNN-LSTM model that targets both spatial and temporal biases, which has potential to further improve the accuracy of the forecasted precipitation and shows the usage of machine learning as a valuable post-processing tool for numerical weather prediction.