欢迎访问《图学学报》 分享到：
图学学报

图学学报 ›› 2022, Vol. 43 ›› Issue (4): 677-684.DOI: 10.11996/JG.j.2095-302X.2022040677

• 计算机图形学与虚拟现实 • 上一篇    下一篇

结合残差时空注意力机制的海面温度预测算法

  

  1. 1. 上海海洋大学信息学院，上海 201306；
    2．国家海洋局东海标准计量中心，上海 201306
  • 出版日期:2022-08-31 发布日期:2022-08-15
  • 通讯作者: 贺琪(1979)，女，副教授，博士。主要研究方向为海洋大数据分析、大数据存储、工作流与业务流程管理、服务计算等
  • 基金资助:
    国家自然科学基金项目(61972240)；国家自然科学基金青年项目(41906179)；上海市科委部分地方高校能力建设项目(20050501900)

Sea surface temperature prediction algorithm combined with residual spatial-temporal attention mechanism

  1. 1. Department of Information Technology, Shanghai Ocean University, Shanghai 201306, China;
    2. East China Sea Standard Metrology Center, State Oceanic Administration, Shanghai 201306, China
  • Online:2022-08-31 Published:2022-08-15
  • Contact: HE Qi (1979), associate professor, Ph.D. Her main research interests cover ocean big data analysis, big data storage, workflow and business process management, and service computing, etc
  • Supported by:
    National Natural Science Foundation of China (61972240); Youth Project of National Natural Science Foundation of China (41906179);
    Capacity Building Project of Some Local Universities of Shanghai Science and Technology Commission (20050501900)

PDF (PC)

82

摘要：

海面温度(SST)与全球气候变化、海洋灾害、海洋生态系统密切相关，因此准确地预测 SST 是一个重要课题。现有区域型 SST 预测方法将 SST 时间序列处理为二维矩阵序列并作为模型输入，每个矩阵对应着特定时刻的区域 SST，通过提取时空特征来实现其预测，但未充分考虑不同时空特征在时间维度和空间维度上对 SST 影响的不均衡性，限制了预测精度地提高。为了解决该问题，提出了一种结合时间注意力机制和空间注意力机制的区域 SST 预测方法(CRA-ConvLSTM)，使得模型动态关注不同时刻的时间特征和区域内不同点的空间特征，赋予不同的影响权重，进而提高 SST 预测精度。具体来说，首先将输入的区域 SST 时间序列通过卷积神经网络(CNN）编码为多层特征向量，提取局部特征；然后构建了残差时间注意力模块，自适应地学习不同时刻的注意力权重，提取时间维度上的关键特征，并设计了残差空间注意力模块，提取区域内不同点在空间维度上的关键特征，此外，将注意力机制结合残差结构避免了网络中信息量过少导致的性能下降问题；最后通过卷积长短时记忆神经网络(ConvLSTM)将特征向量映射为 SST 预测结果。实验结果显示，该模型的均方根误差(RMSE)和预测精度(PACC)分别达到了 0.19 和 99.43%，均优于其他方法，有效提高了 SST 的预测精度。

关键词: 时间序列, 海面温度预测, 时空特征, 注意力机制, 残差结构

Abstract:

Sea surface temperature (SST) is closely related to global climate change, ocean disasters, and ocean ecosystems, so the accurate prediction of SST is an important topic. The existing regional SST prediction methods treat the time series of SST data as a series of matrixes, each corresponding to the regional SST at a particular time. The spatial and temporal features are extracted from the matrix series for later SST prediction. However, the existing SST prediction methods fail to fully consider the imbalanced influence of temporal and spatial features on the SST, leading to the neglection of some key information and limiting the improvement of prediction accuracy. To address this problem, we proposed a regional SST prediction method (CRA-ConvLSTM) combining temporal attention mechanism and spatial attention mechanism. This enabled the model to dynamically assign different influence weights to the temporal features at different times and spatial features at different locations, thereby improving the accuracy of SST prediction. Specifically, the input regional SST time series was first encoded into multi-layer feature vectors by a convolutional neural network (CNN), and local features were extracted. Then, the residual time attention module was constructed to learn the attention weight at different moments adaptively, and the key features of the time dimension were extracted. The residual spatial attention module was designed to extract the key features of different points in the region in terms of the spatial dimension. In addition, the attention mechanism combined with the residual structure can avoid performance degradation caused by information reduction in the network. Experimental results show that the proposed model could achieve 0.19 and 99.43% respectively in terms of the root mean square error (RMSE) and prediction accuracy (PACC), which is superior to other methods and effectively improves the prediction accuracy of SST.

Key words:

中图分类号: