热带海洋学报 ›› 2024, Vol. 43 ›› Issue (1): 40-47.doi: 10.11978/2023037CSTR: 32234.14.2023037

• 海洋水文学 • 上一篇    下一篇

卷积神经网络在近岸表层海温预报中的应用

翁少佳(), 蔡锦海, 庞运禧, 罗荣真()   

  1. 自然资源部汕头海洋中心, 广东 汕尾 516600
  • 收稿日期:2023-03-23 修回日期:2023-04-23 出版日期:2024-01-10 发布日期:2024-01-19
  • 作者简介:

    翁少佳(1990—), 男, 广东省潮州市人, 硕士研究生, 从事海洋观测预报研究。email:

  • 基金资助:
    广东省平台基地及科技基础条件建设项目(2021B1212050025)

Application of convolutional neural network to sea surface temperature prediction in the coastal waters

WENG Shaojia(), CAI Jinhai, PANG Yunxi, LUO Rongzhen()   

  1. Shantou Marine Center, Ministry of Natural Resources, Shanwei 516600, China;
  • Received:2023-03-23 Revised:2023-04-23 Online:2024-01-10 Published:2024-01-19
  • Supported by:
    Science and Technology Plan Projects of Guangdong Province(2021B1212050025)

摘要:

针对数值预报和人工经验预报在近岸定点表层海温(sea surface temperature, SST)预报中预报准确度不高, 将近岸台站定点SST预报转换为多元时间序列预测任务, 应用卷积神经网络(convolutional neural networks, CNN)构建近岸台站定点SST时间序列变化模型, 对近岸台站每日最高海温、最低海温、平均海温进行预报, 并与人工经验方法和长短期记忆网络(long short-term memory, LSTM)方法进行对比试验。结果显示, 在测试数据中相比人工经验预报, CNN方法全年日最高海温预报平均绝对误差(mean absolute error, MAE)为0.36℃, 平均下降0.14℃, 均方根误差(root mean squared error, RMSE)为0.49℃, 平均下降0.21℃, 日最低海温预报MAE为0.36℃, 平均下降0.17℃, RMSE为0.63℃, 平均下降0.24℃, 日平均海温预报MAE为0.30℃, RMSE为0.47℃, 预报性能和LSTM模型预报性能相当。研究表明CNN应用于近岸SST预报具有可行性, 能够有效地提高SST预报准确度, 并且预报效果可以媲美LSTM。

关键词: 表层海温, 近岸台站, 多元时间序列, 卷积神经网络

Abstract:

Concerning the low sea surface temperature (SST) prediction accuracy of numerical modeling and empirical methods in near-shore stations, we consider sea surface temperature prediction as forecasting of multivariate time series data, construct the sea surface temperature time series model of near-shore stations by convolutional neural network (CNN) to predict the maximum, minimum and mean sea surface temperature for the next day, and compare CNN model with empirical forecast method and long short-term memory (LSTM) model through experiment. The experimental results show that compared with empirical forecast method, the mean absolute error (MSE) of CNN model on daily maximum SST forecast drops 0.14℃ to 0.36℃, root mean squared error (RMSE) drops 0.21℃ to 0.49℃, the MSE of CNN model on daily minimum SST forecast drops 0.17℃ to 0.36℃, RMSE drops 0.24℃ to 0.63℃, the MSE of CNN model on daily mean SST forecast is 0.30℃, RMSE is 0.47℃, its forecast performance is as good as LSTM model in the testing set. It shows that the application of CNN to SST modeling is feasible, improve the accuracy of sea surface temperature prediction which can compare favorably with LSTM model.

Key words: sea surface temperature, near-shore stations, multivariate time series, CNN