用于短临降水预报的多尺度注意力编码-动态解码网络  

Multiscale attentional coding-dynamic decoding network for short-range precipitation nowcasting

作　　者：杜先君[1] 郭航飞 程生毅 Du Xianjun;Guo Hangfei;Cheng Shengyi(College of Electrical and Information Engineering,Lanzhou University of Technology,Lanzhou 730050,China)

机构地区：[1]兰州理工大学电气工程与信息工程学院,甘肃兰州730050

出　　处：《海洋学报》2024年第12期122-134,共13页

基　　金：国家自然科学基金(62241307);甘肃省科技计划项目(22YF7FA166、24JRRA173、24CXGA050);兰州市科技计划项目(2024-3-47)。

摘　　要：短临降水预报是气象学和水文学中的重要任务之一,但在现有深度学习方法中,其预测结果模糊不清,并且累计误差大。为了克服这些预测方法中存在累计误差的局限性,以及预测序列结果模糊不清的问题,本文构建了一种基于多尺度注意力编码-动态解码网络(Multi-scale Attention EncodingDynamic Decoding Network, MAEDDN)的短临降水预报方法,通过学习输入数据的时空特征来预测未来的降水情况。为了得到更多输入序列的特征信息,在编码过程中,使用带有空间及通道注意力的卷积块进行编码,并增加多尺度融合模块解决降水分布中小尺度与大尺度信息无法同时捕获的问题;增强预测序列的清晰度,需要模型更好地理解降水过程,因此在解码过程中,针对短临降水过程伴随的生成与消散过程,提出了一种动态解码网络,通过学习输入过去数据的强度分布及变化趋势对解码过程进行灵活地筛选。使用公开数据集SEVIR的降水数据进行实验,并与现有最好模型进行对比,实验结果表明:(1)MAEDDN提升了在高强度降水区域的预测能力。(2)MAEDDN预测的图像序列清晰度显著优于其他模型。构建的多尺度注意力编码能够更好地捕捉气象数据的复杂关系;动态解码能够根据不同的情况自适应地选择解码过程,提供更准确的预测结果。Short-term precipitation nowcasting is a critical task in both meteorology and hydrology.However,current deep learning methods often yield ambiguous prediction results and exhibit significant cumulative errors.To address the limitations associated with these predictive methods,particularly the challenges of cumulative error and lack of clarity in prediction sequences,we propose a novel approach based on a Multi-scale Attention Encoding-Dynamic Decoding Network(MAEDDN)for short-term precipitation nowcasting.This method leverages the learning of spatiotemporal features from input data to accurately predict future precipitation scenarios.To obtain richer feature information from the input sequences,the encoding process employs convolutional blocks with spatial and channel attention for encoding.And a multi-scale fusion module is introduced to address the challenge of capturing both small-scale and large-scale information in precipitation distribution simultaneously.To enhance the clarity of the predicted sequences,the model needs to better understand the precipitation process.Therefore,in the decoding process,a dynamic decoding network is proposed in response to the generation and dissipation processes accompanying short-term precipitation.This network flexibly filters the decoding process by learning the intensity distribution and change trends of past input data.Experiments are conducted by using the precipitation data from the open-source SEVIR dataset,and comparisons are made with the best methods reported so far.The experimental results reveal that:(1)MAEDDN enhances the forecasting capability in areas with high-intensity precipitation,and(2)The clarity of the predicted image sequences by MAEDDN is significantly better than that of other models.The constructed multi-scale attention encoding captures the complex relationships in meteorological data more effectively,while the dynamic decoding adapts the decoding process based on different scenarios,resulting in more accurate prediction outcomes.

关 键 词：短临预报 动态解码网络 时空特征 多尺度融合 

分 类 号：P457.6[天文地球—大气科学及气象学]