一种改进的循环神经网络雷达图像外推算法

An Improved Algorithm of Radar Image Extrapolation Based on Recurrent Neural Network

临近强降水预报目的是预测未来两小时内局地降水强度的分布,准确的外推雷达图像可以为临近强降水预报提供准确的时空参考数据。近两年循环神经网络模型应用于天气雷达回波图象外推得到了较好的结果。本文基于分析现有ConvLSTM和TrajGRU模型的基础上,从输入雷达数据层数和修改模型损失函数两个方面对循环神经网络外推模型进行改进,并对业务上的雷达图象序列和竞赛雷达图象序列进行试验。试验结果表明,改进的外推模型能更好地捕捉时空相关性,具有更精确的外推效果。

The purpose of precipitation nowcasting is to predict the distribution of local precipitation intensity within the coming two hours,and accurate extrapolation radar images can provide accurate spacetime data reference for nowcasting.The application of Recurrent Neural Network(RNN)to meteorological radar image extrapolation has brought better results in recent two years.Based on the analysis of convLSTM and TrajLSTM,the extrapolation model is improved from two aspects:the number of layers of radar data and the loss function,and the experiment is carried out with Chengdu and Yueyang radar data and an open competition data set.The experimental results show that the improved model can better capture the space-time correlation and keep more image details.