基于注意力机制的HDC-BiGRU海表温度预测模型

海表温度（sea surface temperature，SST）是研究全球气候变化的重要地球物理参数，SST的精确预测对全球气候变化、海洋环境和渔业发展具有重要意义。为了提高SST的预测精度，基于时空特征的提取方法，本文提出具有注意力机制的HDC-BiGRU混合模型（HDC-BiGRU-AT，由编码器和解码器构成），可以预测7天的SST。在模型编码阶段，混合空洞卷积（hybrid dilated convolution，HDC）能够提取SST的空间特征，双向门控循环神经网络（bidirectional gated recurrentunit，BiGRU）能够捕获SST的时序特征。通过加入注意力机制，对输出信息分配不同的权重（重要信息分配更高的权重系数），进而实现信息编码，在解码阶段可以提高模型的预测精度。选取我国东海和南海海域的二维SST数据进行建模，实验结果表明，HDC-BiGRU-AT模型的误差指标值均低于已有的方法，充分验证了所提方法的可行性、有效性。

Prediction model for sea surface temperature based on HDC-BiGRU with attention mechanism

Sea surface temperature (SST) is an important geophysical parameter to study global climate change. Accurate prediction of SST is of great significance to global climate change, marine environment and fishery development. In order to improve prediction accuracy of SST, the HDC-BiGRU hybrid model with attention mechanism (HDC-BiGRU-AT) is proposed by extracting the temporal and spatial characteristics to predict the sea surface temperature in the next seven days. The model consists of the encoder and the decoder. In the encoding phase, hybrid dilated convolution (HDC) can extract spatial characteristics of SST, and bidirectional gated recurrent neural network (BiGRU) can capture temporal characteristics of SST. By adding attention mechanism to assign different weights to the output information and assign higher weight coefficients to the important information, the information encoding is realized and the prediction accuracy of the model is improved in the decoding stage. Two-dimensional SST data selected from the East and South China Seas are adopted for modeling. The experimental results show that the values of error indices on the HDC-BiGRU-AT model are lower than those of the existing models, which verifies the feasibility and effectiveness of the proposed method.