基于RGCN-SA算法的海上浮标观测数据插补  

Maritime buoy observation data Imputation based on RGCN-SA algorithm

作　　者：彭德东 梁建峰[1] 崔学荣[2] 岳心阳[1] PENG Dedong;LIANG Jianfeng;CUI Xuerong;YUE Xinyang(National Marine Data and Information Service,Tianjin 300171,China;College of Oceanography and Space Informatics,China University of Petroleum(East China),Qingdao 266580,China)

机构地区：[1]国家海洋信息中心,天津300171 [2]中国石油大学(华东)海洋与空间信息学院,山东青岛266580

出　　处：《海洋预报》2024年第5期77-88,共12页Marine Forecasts

基　　金：国家重点研发计划(2021YFC3101600)。

摘　　要：针对海洋观测数据的缺失问题,提出一种基于图卷积(GCN)和自注意力机制(SA)的残差网络插补模型(RGCN-SA),该模型由自注意力机制与图卷积构建,利用自注意力机制提取观测数据的时间依赖特征,通过图卷积获取不同位置浮标的空间依赖特征,并添加残差结构提高模型学习能力,结合自监督训练方式对模型进行训练,得到最终的海洋浮标数据插补模型。通过对比实验,证明该模型通过训练后能够有效获取浮标观测数据的时间与空间的关联特征,取得了比其他方法更好的插补效果。通过消融实验,证明模型的各个模块的有效性。In this paper,a residual network imputation model based on graph convolution network(GCN)and self-attention mechanism(RGCN-SA)is proposed to solve the observational data missing problem.The model is constructed on self-attention mechanism and graph convolution.The self-attention mechanism is used to extract the time-dependent features of observational data,and the space-dependent features of buoys at different positions are obtained through graph convolution.Combined with the self-supervised training method,the model is trained and the final ocean data imputation model is obtained.Through comparative experiments,it is proved that the model can effectively obtain the temporal and spatial correlation features of buoy observations after training,and obtaina better imputation effect than other methods.The effectiveness of each module of the model is proved by the ablation experiment.

关 键 词：自注意力机制 图卷积网络 插补 浮标数据 

分 类 号：P731.31[天文地球—海洋科学]