基于ConvGRU深度学习网络模型的海表面温度预测

doi:10.16535/j.cnki.dlhyxb.2021-137

PDF(6790 KB)

大连海洋大学学报 ›› 2022, Vol. 37 ›› Issue (3) : 531. DOI: 10.16535/j.cnki.dlhyxb.2021-137

基于ConvGRU深度学习网络模型的海表面温度预测

张雪薇，韩震*

作者信息 +

Prediction of sea surface temperature based on ConvGRU deep learning network model

ZHANG Xuewei,HAN Zhen*

Author information +

文章历史 +

摘要

为解决传统时间序列网络模型的时间性与空间性结合不足和批量处理海洋环境要素大数据能力不足的问题，采用循环神经网络(RNN)扩展算法和卷积神经网络(CNN)相结合的ConvGRU深度学习神经网络模型以有效体现时空特征，利用样本生成器处理数据以有效处理时间序列遥感数据的批量性问题，并使用1999—2019年海表面温度网格数据(OISST产品)，对2020年西北太平洋部分海域进行了海表面温度预测分析。结果表明：ConvGRU模型训练集的均方根误差和准确率分别为0.044 9 ℃和99.69%，验证集的均方根误差和准确率分别为0.045 2 ℃和99.64%；使用建立的ConvGRU模型对2020年海表面温度数据进行了预测，测试集的均方根误差和准确率分别为0.047 8 ℃和99.60%，海表面温度预测值的平均绝对误差和预测精度的平均值分别为0.379 3、97.31%。研究表明，本文中建立的ConvGRU模型可以较好地预测海表面温度的变化趋势，这为海表面温度神经网络预测模型提供了一种可行性方法。

Abstract

In order to solve the problems of insufficient combination of temporality and spatiality of traditional time series network models and insufficient ability of batch processing large data of marine environmental elements,sea surface temperature (SST) prediction and analysis were carried out using the SST grid data from 1999 to 2019 in part of the Northwest Pacific Ocean (OISST Products) by ConvGRU deep learning neural network model,that is,combination of the extended algorithm of recurrent neural network (RNN) and convolutional neural network (CNN),and by sample generator to process data,which can effectively deal with the batch problem of time series remote sensing data.The results showed that the ConvGRU model training set had root mean square error of 0.044 9 ℃ and accuracy of 99.69%,and that the validation set had root mean square error of 0.045 2 ℃ and accuracy of 99.64%.Finally,the model established in this paper was used to predict the sea surface temperature data in 2020,with the root mean square error of 0.047 8 ℃ and accuracy of 99.60% in the test set.The average mean absolute deviation of SST prediction was shown to be about 0.379 3 ℃,and the average prediction accuracy to be 97.31%.The findings indicate that the ConvGRU neural network model can predict the trend of sea surface temperature well,which provides a feasible method for SST neural network prediction model.

导出引用

张雪薇, 韩震. 基于ConvGRU深度学习网络模型的海表面温度预测[J]. 大连海洋大学学报, 2022, 37(3): 531 https://doi.org/10.16535/j.cnki.dlhyxb.2021-137

ZHANG Xuewei, HAN Zhen. Prediction of sea surface temperature based on ConvGRU deep learning network model[J]. Journal of Dalian Fisheries University, 2022, 37(3): 531 https://doi.org/10.16535/j.cnki.dlhyxb.2021-137

中图分类号： S 915 TP 79