Traditional linear models are insufficient for capturing the complex dynamics of sea level changes. This paper aims to predict sea level time series using neural network models adapted for nonlinear data. Currently, few researchers use a combination of convolutional neural networks and bidirectional gated recurrent units (CNN-BiGRU) optimized with hyperparameter tuning for sea level prediction. There is also a lack of detailed discussion on the randomness of neural network initialization in prediction domains. Therefore, this study employs the bayesian optimization algorithm (BO) to optimize the CNN-BiGRU model, resulting in the BO-CNN-BiGRU model. Experiments initially compared the BO-CNN-BiGRU model with five other models using data from ten tidal stations in the US, showing that the model outperformed the others. To address initialization randomness, we used ten random seeds for statistical analysis, which demonstrated that the BO-CNN-BiGRU model performed well in terms of predictive performance and robustness. Finally, the BO-CNN-BiGRU model was applied to satellite altimetry grid data from the Bohai and Yellow Seas in China, yielding a linear trend of 3.92 ± 0.63 mm/a from 1993 to 2023, consistent with the China Sea Level Bulletin, further validating the model's effectiveness. This model can be used to predict regional sea level change.
Read full abstract