Short-Term Precipitation Prediction for Contiguous United States Using Deep Learning

Accurate short-term weather prediction, essential for many aspects of life, relies mainly on forecasts from numerical weather models. Here, we report results supporting strongly deep learning as a viable, alternative approach. A 3D convolutional neural network, which uses a single frame of meteorology fields as input to predict the precipitation spatial distribution, is developed based on 39-years (1980-2018) data of meteorology and daily precipitation over the contiguous United States. Results show that the trained network outperforms the state-of-the-art weather models in predicting daily total precipitation, and the superiority of the network extends to forecast leads up to 5 days. Combining the network predictions with the weather-model forecasts significantly improves the accuracy of model forecasts, especially for heavy-precipitation events. Furthermore, the millisecond-scale inference time of the network facilitates large ensemble predictions for extra accuracy improvement. These results demonstrate the promising prospects of deep learning in short-term weather predictions.