Deep Learning Prediction of Incoming Rainfalls: An Operational Service for the City of Beijing China

We developed a deep learning prototype for the purpose of rainfall prediction for the city of Beijing China. It predicts rainfall dynamics in the next two hours with spatial reso-lution of 1km and temporal resolution of 6 minutes. That is a magnitude better than traditional weather forecasting. The computation time is less than 10 seconds, several magnitudes of less computation time than traditional forecasting. This deep learning network combines the strengths of known structures such as U-Net, ResNet, Sqeeze-and-Excitation, and the spatial Attention mod-ule. We rely solely on full convolutional layers instead of RNN layers as used in other weather prediction ef-forts. Meteorologically assessment metrics suggest better performance by the FCN approach. The prototype is now up and running in anticipation for the summer rain season of 2019. This might be the first such service employed by a capital weather service.