CAST-LSTM: A Spatio-Temporal LSTM Model for Radar Echo Extrapolation
The forecast results of radar echo extrapolation algorithm based on recurrent neural network are gradually blurred and distorted with time, and it is difficult to forecast the high echo area. To solve the above problems, this paper proposes a spatio-temporal long short-term memory network model based on context fusion and attention mechanism. The method fully extracts the short-term context information of different scales of radar image through the context fusion module. The attention module broadens the time perception domain of the prediction unit, so that the model perceives more time dynamics. Taking the weather radar data of Jiangsu Province from April to September in 2019-2021 as a sample, the spatio-temporal long short-term memory network based on context fusion and attention mechanism achieves better prediction performance through experimental comparison and analysis. The Critical Success Index (CSI) and Heidke Skill Score (HSS) reached 0.7611, 0.5326, 0.2369 and 0.7335, 0.5735, 0.3075, respectively, under the conditions of 60 minutes of extrapolation and thresholds of 10, 20 and 40 dBZ, which effectively improved the prediction accuracy.