摘要
为了实现对中国东部地区极端强降水的临近预报、预警,基于具有物理约束功能的PhyDNet构建了融合雷达反射率因子和分钟级降水观测资料的融合降水临近预报模型PhyDNet-RP,预测江苏省及其上游地区未来3 h降水量,对比和探究了PhyDNet-RP、INCA(交叉相关外推+中尺度模式融合)、PhyDNet-P(仅包含降水资料)和UNet-RP(融合因子与PhyDNet-RP相同,但采用UNet模型)4种临近预报方法及对强降水增强过程的预测能力。结果表明:(1)与INCA相比,深度学习方法能更好地体现强降水增强过程的发展和演变,(2)对比PhyDNet-P和PhyDNet-RP模拟结果发现,在深度学习模型输入资料中增加雷达反射率因子可以更好地再现强降水区的形状和移动特征,(3)UNet-RP能够再现降水区的形状和移动,但不能定量降水强度。4种方法中,PhyDNet-RP预报效果最优,说明在模型输入资料中叠加具有不同功能属性的通道因子对预测效果具有正贡献,为深度学习的可解释性提供了一定支撑。
A deep learning-based precipitation nowcast model(PhyDNet-RP)over Jiangsu area is established using a physically-constrained convolutional neural network PhyDNet with both radar reflectivity and ground precipitation observations as inputs.PhyDNet-RP is then compared with INCA(blend TREC with mesoscale numerical model),PhyDNet-P(only precipitation is used as input)and UNet-RP(both precipitation and radar data are used as input for convolutional neural network UNet),especially under convection enhancement scenarios.Results are as follows:(1)The deep learning-based nowcast model performs better than INCA on the forecast of precipitation growth and decay.(2)The comparison of PhyDNet-P and PhyDNet-RP reveals that adding radar reflectivity as a second input factor can improve the forecast of movement and form of strong precipitation.(3)UNet-RP can to some extent reflect the movement and form of precipitation,but it fails to capture the intensity.It is concluded that PhyDNet-RP shows an overall advantage over INCA,PhyDNet-P,and UNet-RP,indicating that precipitation nowcast obviously benefits from employing different factors as input for deep learning model.
作者
庄潇然
郑玉
王亚强
康志明
闵锦忠
张文华
李杨
ZHUANG Xiaoran;ZHENG Yu;WANG Yaqiang;KANG Zhiming;MIN Jinzhong;ZHANG Wenhua;Li Yang(Jiangsu Meteorological Observatory,Nanjing 210041,China;Key Laboratory of Transportation Meteorology,Nanjing Joint Institute for Atmospheric Sciences,Nanjing 210041,China;Chinese Academy of Meteorological Sciences,Beijing 100081,China;Nanjing University of Information Science and Technology,Nanjing 210044,China)
出处
《气象学报》
CAS
CSCD
北大核心
2023年第2期286-303,共18页
Acta Meteorologica Sinica
基金
国家自然科学基金项目(42105008、U2142210)
中国气象局重点创新团队项目(CMA2022ZD04)
南京气象科技创新研究院北极阁开放基金项目(BJG202103)
江苏省气象局科研项目(KQ202211)。
关键词
降水临近预报
深度学习
天气雷达
雨量站
物理约束模型
Precipitation nowcast
Deep learning
Weather radar
Rainfall station
Physically-constrained model
