王善昊,胡志群. 2024. 基于ConvLSTM融合RMAPS-NOW数据的雷达回波外推研究[J]. 气象学报, (0):-, doi:[doi]
基于ConvLSTM融合RMAPS-NOW数据的雷达回波外推研究
Research on radar echo extrapolation based on ConvLSTM fusion of RMAPS-NOW data
投稿时间:2023-09-20  修订日期:2024-01-16
DOI:
中文关键词:  雷达回波外推,深度学习,RMAPS-NOW,MR-ConvLSTM网络架构,自定义损失函数
英文关键词:Radar echo extrapolation, deep learning, RMAPS-NOW, MR-ConvLSTM network, self-defined loss function
基金项目:
作者单位邮编
王善昊 成都信息工程大学电子工程学院 610225
胡志群* 中国气象科学研究院灾害天气国家重点实验室 100081
摘要点击次数: 12
全文下载次数: 5
中文摘要:
      [目的和资料方法]雷达回波外推是临近预报、人工影响天气作业及效果评估的主要参考依据之一,快速准确的回波外推技术,一直是雷达气象领域的研究热点。近年来,基于深度学习的时空序列预测模型在雷达回波外推中得到了广泛的应用。然而,这些外推网络架构的输入大多用16级色标伪彩色雷达回波强度产品转化而来的灰度图,丢失许多回波细节,并且随着外推时间增加,误差不可避免增大。回波的生消移动演变与天气背景紧密相关,因此,本文将北京城市气象研究院研发的新一代快速更新多尺度资料分析和预报系统的临近数值预报子系统(RMAPS-NOW)初始零场的部分物理量产品融入华北雷达拼图原始数据,构建多个雷达单元(Radar cells),并将这些Radar cells作为输入,基于卷积长短期记忆网络(ConvLSTM),设计了一个多通道雷达回波外推网络架构(MR-ConvLSTM)。另外,考虑到卷积算法的平滑性,构建了自定义损失函数,增加回波强度的时空权重进行时空衰减订正。选取114?E—115.4?E,40.65?N—41.65?N范围内2018—2021年的6—9月共13000组华北雷达组合反射率因子拼图,及RMAPS-NOW初始零场数据,其中80%共10400组为训练集,20%共2600组为测试集。引入的物理量包括多个高度层的u,v风(1350m),相对湿度(150m),水平散度(1350m)等,基于ConvLSTM,以及MR-ConvLSTM加自定义损失函数,分别训练得到5个雷达回波外推模型。[结果与结论]采用临界成功指数(CSI)、命中率(POD)、虚警率(FAR)作为评价指标,利用测试集对所有模型进行评估。基于引入物理量的MR-ConvLSTM训练得到的模型在20、30、35dBz反射率阈值下,比未引入物理量的基于ConvLSTM的外推模型CSI值平均高出4.67%、13.8%、5.98%,POD值平均高出3.1%、7.68%、8.38%,FAR值平均低出6.37%、8.54%、10.17%,同时引入三种物理量(rh、uwind、vwind)的外推模型在不同阈值的各项指标中综合表现为最好,其CSI、POD值在三种不同阈值下比未引入物理量模型平均高出16.01%、13.38%,FAR值平均低出14.88%。从模型应用的个例可视化也可以看出,引入物理量后有效提升了雷达回波外推的准确性,证明了基于MR-ConvLSTM架构训练的雷达回波外推模型有较强的泛化能力。
英文摘要:
      Radar echo extrapolation is one of the main reference criteria for nowcasting, weather modification operations and its effectiveness evaluation. Therefore, rapid and accurate echo extrapolation technology is always a research hotspot in the field of radar meteorology. In recent years, deep learning-based spatiotemporal sequence prediction models have been widely applied in radar echo extrapolation. However, most of the inputs of these extrapolation network are only grayscale images converted from the radar intensity products shown with 16 level color code pseudo color, so some echo details are lost, and the error inevitably increase with the extrapolation time. The initiation, disappearance, movement, and evolution of echoes are closely related to the weather background. Based on this consideration, some physical products in the initial zero field of the rapid-refresh multi-scale analysis and prediction system-nowcasting (RMAPS-NOW) developed by the institute of Urban Meteorology, CMA, are integrated with the raw data of the North China radar mosaic to construct multiple radar cells. Based on convolutional long short-term memory (ConvLSTM) network, a multi-channel radar echo extrapolation (MR-ConvLSTM) network is designed by using the radar cells as inputs. In addition, considering the smoothness of the convolutional algorithm, a self-defined loss function is designed to increase the spatiotemporal weight of the echo intensity for spatiotemporal attenuation correction. A total of 13,000 samples of radar mosaic and RMAPS-NOW data from June to September, during 2018 to 2021 within 114?E to 115.4?E,40.65?N to 41.65?N are selected. In which 80%, i.e. 10,400 samples are as training dataset, and 20%, i.e. 2,600 as test dataset. The introduced physical products include u, v wind (1350m), relative humidity (150m), and horizontal divergence (1350m), etc. at multiple altitude levels. And then, based on ConvLSTM and MR-ConvLSTM with the self-defined loss function, 5 extrapolation models are trained, respectively.Using critical success index (CSI), hit rate (POD), and false alarm rate (FAR) as evaluation indicators, the models are evaluated by the test dataset. At the values of reflectivity threshold 20, 30, and 35 dBZ, the average values of CSI calculated by the MR-ConvLSTM-based and the self-defined function models integrated physical products are 4.67%, 13.8%、5.98% higher, and the average POD are 3.1%、7.68%、8.38% higher, the average FAR are 6.37%, 8.54%, 10.17% lower than those by the ConvLSTM-based model without integrating physical products, respectively. The model introduced with three physical products (rh, uwind, vwind) performs the best in all the indicators, which average CSI, POD are 16.01%, 13.38% higher, and FAR is 14.88% lower than those without physical products, respectively. From the cases visualization of model application, it can also be seen that the introduction of physical quantities effectively improves the accuracy of radar echo extrapolation It proves that the MR-ConvLSTM models and self-defined loss function have robust generalization ability.
查看全文   查看/发表评论  下载PDF阅读器