卷积门控循环单元神经网络与光流法在临近预报中的适用性研究

Applicability of Convolutional Gated Recurrent Unit Neural Network and Optical Flow Method in Nowcasting

深度学习目前在临近预报的雷达外推应用中发展迅速,对其适用性的客观评估是业务应用的重要前提。利用粤港澳大湾区雷达回波开放数据集,基于回波图像形态、命中率、虚警率以及技巧评分,对比评估了卷积门控循环单元神经网络(ConvGRU)与基于半拉格朗日平流方案的快速稠密光流法(OF)在未来120 min雷达回波外推中的效果,结果表明ConvGRU与OF虽外推效果良好,但均不适用于外推回波的生成、加强以及局地分散性特征,且外推效果显著受到天气过程种类的影响。其中,ConvGRU可能更适用于外推分布范围适中且运动简单的回波的主体位置,但回波的强度不稳定且形态模糊,其无法表征回波的精细化运动规律与演化特征;OF更适用于外推50 dBz以上的强回波,且回波结构更优,但在回波缺测区与强度少变区内的外推易存在很大的分布偏差。深度学习的数据集样本数是决定模型效果的最重要原因之一,但样本数实际未能全面覆盖各类天气过程,总体仍偏少,预报业务需进一步扩增。

Deep learning has been developed at an unprecedend speed in radar extrapolation of forecasting,so objective assessment of its applicability is an important prerequisites for operational applications.By utilizing the radar echo open data set of Guangdong-Hong Kong-Macao Greater Bay Area,the performances of 120 min radar echo extrapolation by convolutional gated recurrent unit neural network(ConvGRU)and fast dense optical flow(OF)methods based on semi-lagrangian advection scheme have been compared and evaluated based on the echo morphology,probability of detection(POD),false alarm rate(FAR)and threat score(TS).The results show that although the two methods both have effective extrapolation performance,they are not applicable to extrapolate the echo generation,enhancement and locally dispersed.The extrapolation effects are significantly affected by the type of weather processes.ConvGRU,poor in characterizing the refined motion pattern and evolution of the echoes,may be more suitable for extrapolating the main location of echoes with moderate distribution and simple motion,but the extrapolated echoes are unstable in intensity with blurred forms.OF is more suitable for extrapolating the strong echoes above 50 dBz and has better echoes structure but it is prone to large distribution deviations in the region with missing or stable echoes.The number of samples,one of the most important elements to determine the model effectiveness,is still not enough in this paper,which needs to be further expanded for covering all kinds of weather processes comprehensively in actual forecast operation.