基于卷积神经网络的雷达回波外推方法

摘要

Extrapolation technique of weather radar echo possesses a widely application prospects in short-term nowcast.The traditional methods of radar echo extrapolation are difficult to obtain long limitation period and have low utilization rate of radar data.This problem is researched from deep learning perspective in this paper,and a new model named Dynamic Convolutional Neural Network based on Input (DCNN-I) was proposed.According to the strong correlation between weather radar echo images at adjacent times,dynamic sub-network and probability prediction layer were added,and a function was created that maped the convolution kernels to the input,through which the convolution kernels could be updated based on the input weather radar echo images during the testing.In the experiments of radar data from Nanjing,Hangzhuo and Xiamen,this method achieved higher accuracy of prediction images compared with traditional methods,and extended the limitation period of exploration effectively.%雷达回波外推技术目前被广泛应用于临近预报中.针对传统雷达回波外推方法存在外推时效较短,对雷达资料数据利用率不高的问题,采取深度学习的方法,提出了一种基于输入的动态卷积神经网络(DCNN-I)模型.根据相邻时刻的雷达回波图像之间相关性强的特点,该网络模型中增加了动态子网络和概率预测层,建立了卷积核与输入图像的映射关系,使卷积核在网络测试阶段仍然能够根据输入雷达回波图像的不同而变化,增强了预测图像与输入图像之间的关联.以南京、杭州、厦门三地的雷达数据为样本进行实验,实验结果表明,与传统的雷达回波外推方法相比,所提方法能够获得更高的预测图像准确率,并且有效延长外推时效.