| 摘要: |
| 海面温度(sea surface temperature, SST)是影响海洋气候变化的关键因素之一, SST 的精确预测对海洋气象、航海等相关领域具有重要意义。为同时捕获空间和时间相关性, 本文提出了一种融合图卷积(graph convolution, GC)和注意力机制的门控循环单元(gated recurrent unit, GRU)海面温度预测模型(graph convolutional recurrent unit-attention mechanism, GCRU-ATT)。 GC 将海洋表面空间构建成图形的拓扑结构, 有效地挖掘数据特有的空间特征。首先, 将门控循环单元中的矩阵乘法替换为图卷积运算,构成门控图卷积(graph convolutional recurrent unit, GCRU)层; 应用 GCRU 层搭建模型主要结构, 以提取数据的时空信息; 其次, 引入注意力机制为 GCRU 层输出信息分配不同的权重。最终, 使用一个全连接的输出层输出海面温度预测结果。选取东海和渤海海域的 SST 数据建模, 实验结果表明GCRU-ATT 模型鲁棒性好, 且其误差指标值低于已有的方法, 预测精度较高。 |
| 关键词: 海面温度 图卷积 门控循环单元 注意力机制 |
| DOI：10.11759/hykx20230523002 |
| 分类号:P731.3 |
| 基金项目:南方海洋科学与工程广东省实验室 (珠海 )资助项目(SML2020SP007); 国家自然科学基金项目(42192562) |
|
| A GRU–sea surface temperature prediction model integrating graph convolution and an attention mechanism |
|
WANG Lina1,2, SONG Yue1, WANG Xudong1, LÜ Luying1, DONG Changming2,3
|
|
1.School of Artificial Intelligence/School of Future Technology, Nanjing University of Information Science and Technology, Nanjing 210044, China;2.Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai), Zhuhai 519080, China;3.School of Marine Sciences, Nanjing University of Information Science and Technology, Nanjing 210044, China
|
| Abstract: |
| Sea surface temperature (SST) is a key factor that affects oceanic climate changes; consequently, the accurate prediction of SST is of great significance in related fields such as oceanic meteorology and navigation. To simultaneously capture spatiotemporal correlations between the SST data, this paper proposes a gated recurrent unit (GRU)–SST prediction model (graph convolutional recurrent unit–attention mechanism, GCRU–ATT) that combines graph convolution (GC) and an attention mechanism. A sea surface space is modeled into a graphical topological structure through GC, which is subsequently used to effectively mine the unique spatial features of the SST data. Initially, in the GRU, matrix multiplication is replaced with GC and GCRU layers are formed. These GCRU layers are then used to build the main structure of the model to extract the spatiotemporal information of the data. Further, an attention mechanism is introduced to assign different weights to the output information of the GCRU layers. Finally, a fully connected output layer is used to output the SST prediction results. To select SST data from the East China Sea and Bohai Sea for modeling, experimental results show that the GCRU–ATT model exhibits superior robustness, smaller error index values, and higher prediction accuracy than existing methods. |
| Key words: sea surface temperature graph convolution gated recurrent unit attention mechanism |
