人民黄河2025，Vol.47Issue(8)：27-31,38,6.DOI:10.3969/j.issn.1000-1379.2025.08.006
基于变分自编码器的临近降水预报技术研究
Research on Nowcasting Precipitation Technology Based on Variational Autoencoder
摘要
Abstract
In order to address the issue that current nowcasting precipitation models were unable to learn nonlinear feature variations of radar echo images in the spatial and temporal dimensions,this paper proposed a precipitation nowcasting method based on a variational autoencoder(VAE).a)A variational autoencoder was built to define a probability distribution function in the latent space for the rebuilding of radar echo images.b)The self-attention mechanism was employed to learn the dependencies of radar echo images in spatial and temporal dimensions.c)The discrete latent space was introduced to capture the complex contextual semantic information of radar echo images.The model perform-ance evaluation experiments were conducted on the SEVIR dataset,with comparative analyses against the two representative precipitation nowcasting models of Simvp and PhyDNet.The results indicate that this model achieves accurate prediction of radar echo images of the next 5 frames,and its prediction accuracy is higher than that of the Simvp and PhyDNet models.The introduction of each module can contribute to the performance improvement of the precipitation nowcasting model.关键词
临近降水预报/深度学习/变分自编码器/自注意力机制Key words
nowcasting of precipitation/deep learning/variational autoencoder/self-attention mechanism分类
天文与地球科学引用本文复制引用
胡明明,尹君逸,郭森,司晓云,陶文彬,郭景涛..基于变分自编码器的临近降水预报技术研究[J].人民黄河,2025,47(8):27-31,38,6.基金项目
河南省科技攻关项目(232102210114) （232102210114）
