关键词: 海面高, 奇异谱分析, 长短期记忆神经网络, 自回归移动平均, 验潮站, 日本海, 预测方法

Abstract:

With the increasing influence of global climate change on sea levels, accurate prediction of sea-surface-height variations is essential for marine environmental monitoring, disaster early warning, and coastal resource management. Using sea-surface-height observations from 28 tide gauge stations along the coast of Japan from 2004 to 2023, this study proposes a prediction framework that integrates Singular Spectrum Analysis (SSA), a Long Short-Term Memory (LSTM) model, and an Autoregressive Moving Average (ARMA) model. The proposed method is applied to predict sea surface height along the Japanese coast and is compared with an SSA+ARMA hybrid approach. The results indicate that the SSA+LSTM+ARMA model outperforms the conventional SSA+ARMA approach in predicting sea surface height for 2019 to 2023. Specifically, the annual Mean Absolute Error (MAE) is reduced by approximately 6%~8% relative to SSA+ARMA. Projections for 2024 to 2025 suggest that sea-surface-height variations will continue to exhibit historical periodicity without significant anomalies, with an overall stable trend. These findings demonstrate that the SSA+LSTM+ARMA model more effectively captures nonlinear dynamics and long-term trends in sea surface height, providing a useful reference for marine management and disaster early warning.

Key words: sea surface height, Singular Spectrum Analysis, Long Short-Term Memory neural network, Autoregressive Moving Average, tide gauge, coast of Japan, prediction method