基于深度学习的高时空分辨率降水临近预报方法

降水临近预报在强对流天气监测预警中具有重要地位,对于防灾减灾至关重要。在气象业务中,主要采用雷达回波外推方法解决高时空分辨率的临近预报问题。针对传统雷达回波外推方法中普遍存在的资料信息利用率不足和预报准确率低的问题,利用上海地区多年的高时空分辨率天气雷达探测资料,基于数据驱动的深度学习方法进行雷达回波外推,提出了一种新的降水临近预报模型——SwinAt-UNet模型。该预报模型通过融合UNet模型和Swin Transformer结构捕捉历史天气雷达探测资料中的短期和长期动态变化特征,可以自适应地学习潜在的雷达回波生消演变规律。此外,为提高模型的泛化能力和预报准确率,引入深度可分离卷积和卷积块注意力模块。结果表明:在不同基本反射率阈值下,SwinAt-UNet模型的预报准确率均高于UNet、SmaAt-UNet、TransUNet和AA-TransUNet模型; 在45 dBZ的基本反射率阈值下,SwinAt-UNet模型临界成功指数提高了13%,同时在预报时效上具有一定的优越性; SwinAt-UNet模型外推图像具有更加清晰的边缘和细节性纹理,对降水范围、移动方向和强度变化的预测更为准确。

Precipitation Nowcasting Method with High Spatio-temporal Resolution Based on Deep Learning

Precipitation nowcasting plays an important role in severe convective weather monitoring and warning, and is very important for disaster prevention and mitigation. In meteorological services, the radar echo extrapolation method is mainly used to solve the nowcasting problem with high spatio-temporal resolution. In order to solve the problems of insufficient utilization of data information and low forecast accuracy in traditional radar echo extrapolation method, the high spatio-temporal resolution weather radar detection data in Shanghai area over many years were used to extrapolate radar echo based on data-driven deep learning method, and a new precipitation nowcasting model(SwinAt-UNet)was proposed. By fusing the UNet model and Swin Transformer structure to capture the short-term and long-term dynamic variation characteristics of historical weather radar detection data, the SwinAt-UNet model could adaptively learn the potential radar echo evolution laws of generation, dissipation, accumulation and deformation. In addition, in order to improve the generalization ability and forecast accuracy of the model, the depthwise-separable convolution and convolutional block attention module(CBAM)were introduced. The results show that the forecast accuracy of SwinAt-UNet model is higher than that of UNet, SmaAt-UNet, TransUNet and AA-TransUNet models under different base reflectivity thresholds; the critical success index of SwinAt-UNet model is increased by 13% at the base reflectivity threshold of 45 dBZ, and the period validity is improved; the image extrapolated by SwinAt-UNet model has clearer edge and detailed texture, and the predictions of precipitation range, moving direction and intensity change are more accurate.