周国良, 周锋. 基于IWOA优化Res-BiGRU深度学习模型的海表温度预测方法[J]. 海洋环境科学, 2024, 43(5): 806-816. DOI: 10.12111/j.mes.2023-x-0126
引用本文: 周国良, 周锋. 基于IWOA优化Res-BiGRU深度学习模型的海表温度预测方法[J]. 海洋环境科学, 2024, 43(5): 806-816. DOI: 10.12111/j.mes.2023-x-0126
ZHOU Guoliang, ZHOU Feng. Sea surface temperature prediction method based on an IWOA optimized Res-BiGRU deep learning model[J]. Chinese Journal of MARINE ENVIRONMENTAL SCIENCE, 2024, 43(5): 806-816. DOI: 10.12111/j.mes.2023-x-0126
Citation: ZHOU Guoliang, ZHOU Feng. Sea surface temperature prediction method based on an IWOA optimized Res-BiGRU deep learning model[J]. Chinese Journal of MARINE ENVIRONMENTAL SCIENCE, 2024, 43(5): 806-816. DOI: 10.12111/j.mes.2023-x-0126

基于IWOA优化Res-BiGRU深度学习模型的海表温度预测方法

Sea surface temperature prediction method based on an IWOA optimized Res-BiGRU deep learning model

  • 摘要: 海表温度(sea surface temperature, SST)是研究海洋环境和气候变化的重点要素之一。现有基于深度学习的SST预测算法存在调参耗时和预测精度偏低的问题,基于此本文提出了一种基于改进鲸鱼优化(improved whale optimization algorithm, IWOA)的残差双向门控循环神经网络(residual bidirectional gated recurrent neural network, Res-BiGRU)的SST预测算法,并选取东海海域的二维月均SST空间分布和一维日均SST时间序列数据验证了算法的可行性和有效性。针对二维月均SST空间分布数据,首先利用Res-BiGRU神经网络提取SST序列信号的空间和时间特征,在解码阶段引入时空注意力机制对提取特征的权重分配进行调整。其次采用IWOA算法对训练参数进行寻优,提高训练参数调整效率和Res-BiGRU模型精度。针对一维日均SST时间序列数据,引入集合经验模态分解法(ensemble empirical mode decomposition, EEMD)对SST序列信号进行预处理,然后利用IWOA-Res-BiGRU进行特征提取和SST预测。实验结果表明,本研究提出的基于IWOA优化Res-BiGRU算法对二维月均SST和一维日均SST的预测均方误差分别可达到0.35 ℃和0.09 ℃,预测结果优于其他模型,可成为SST预测的重要工具。

     

    Abstract: Sea surface temperature (SST) is one of the key elements in the study of marine environment and climate change. The existing SST prediction algorithm based on deep learning has the problems of time-consuming parameter adjustment and low prediction accuracy. Based on this, a SST prediction algorithm based on improved whale optimization (IWOA) residual bidirectional gated recurrent neural network (Res-BiGRU) is proposed. The two-dimensional monthly average SST spatial distribution and one-dimensional SST time series data in the East China Sea are selected to verify the feasibility and effectiveness of the algorithm. For the two-dimensional monthly average SST spatial distribution data, the Res-BiGRU neural network is used to extract the spatial and temporal features of the SST sequence signal, and the time and space attention mechanism is introduced in the decoding stage to adjust the weight distribution of the extracted features. Secondly, the IWOA algorithm is used to optimize the training parameters to improve the efficiency of training parameter adjustment and the accuracy of Res-BiGRU model. Aiming at the one-dimensional daily average SST time series data, the ensemble empirical mode decomposition (EEMD) method is introduced to preprocess the SST sequence signal, and then the IWOA-Res-BiGRU is used for feature extraction and SST prediction. The experimental results show that the proposed Res-BiGRU algorithm based on IWOA optimization can predict the mean square error of two-dimensional monthly mean SST and one-dimensional daily mean SST by 0.35 ℃ and 0.09 ℃, respectively. The prediction results are better than other models and can become one of the important tools for SST prediction.

     

/

返回文章
返回