An Improved Deep Network with Fusion of Physical and Residual Information for Radar Echo Extrapolation

Abstract

Precipitation Nowcasting tasks have important social impact and economic significance in disaster warning, planning, etc. Currently it uses methods based on radar echo extrapolation. Some recent works have introduced deep neural networks into the field of radar echo extrapolation, where radar echo images are arranged in temporal order as inputs to the model. Previous works usually use convolutional neural networks(CNN), recurrent neural networks(RNN) or a combination of both, which do not fully exploit the partial differential equation relationships implied behind the radar echo motion. In this paper, a physical cell is incorporated into the original deep neural network, to add physical constraints to the data-driven model and collaborate with the LSTM variant cell to accurately model the trajectory and appearance of moving objects. The experiments achieve better predictions than the original model on the synthetic moving digital dataset and the realistic radar echo dataset.