基于自注意力机制的深度学习的海洋三维温度场预测  

Three-Dimensional(3-D)ocean temperature field prediction based on deep learning of self-attention mechanism

作　　者：岳伟豪 徐永生[2,3,4] 朱善良 YUE Weihao;XU Yongsheng;ZHU Shaniang(Qingdao university of science and technology,Qingdao 266061,China;The Institute of Oceanology,Chinese Academy of Sciences,Qingdao 266071,China;University of Chinese Academy of Sciences,Beijing 100094,China;Qingdao Marine Science and Technology Center,Qingdao 266000,China)

机构地区：[1]青岛科技大学,山东青岛266061 [2]中国科学院海洋研究所,山东青岛266071 [3]中国科学院大学,北京100094 [4]青岛海洋科技中心,山东青岛266000

出　　处：《海洋预报》2024年第3期22-32,共11页Marine Forecasts

基　　金：崂山实验室科技创新项目(LSKJ202201406);国家自然科学基金(41906027);国家自然科学基金联合基金项目(U22A20587);国家自然科学基金-山东联合基金重点项目(U1406401);中国科学院战略先导计划(XDB42000000)。

摘　　要：目前主要从时空角度出发对海洋三维温度场进行预测,却忽略了相邻位置的海温相关关系。为弥补这一不足,构建一种融合了自注意力记忆模块与卷积式长短时记忆神经网络(ConvLSTM)模型的SA-ConvLSTM三维温度场预测模型,不仅可以从历史三维温度场中提取海温时空特征,还能获取并记忆相邻点位置信息,从而实现对三维温度场时空变化的把握。实验结果表明:相较于ConvLSTM模型,SA-ConvLSTM模型在滑动预测与多步长递归预测实验下的均方根误差和平均绝对误差提升约14%,且整体预测效果均优于基线模型、长短时记忆神经网络模型和ConvLSTM模型。While previous researches on the 3-D ocean temperature field prediction mainly focused on the perspective of spatial and temporal relationship which ignored the relationship of relative location,this article proposes a SA-ConvLSTM 3-D ocean temperature field prediction model which combines self-attention memory module and ConvLSTM.The new model is not only able to extract the spatial-temporal features in historical 3-D ocean temperature fields,but can obtain and record the information of location to learn the laws of seawater in both space and time.The experimental results show that the RMSE and MAE of the SA-ConvLSTM forecasts have approximately increased by 14%in sliding prediction and multi-step recursive prediction,and its overall performance is better than the persistence,LSTM and ConvLSTM model.Our research provides a new idea for the prediction of the 3-D seawater temperature field.

关 键 词：海水温度 三维温度场预测 自注意力记忆机制 SA-ConvLSTM 多步长递归预测 

分 类 号：P731.31[天文地球—海洋科学]