An optimized network model for sea level height prediction integrating OLSDBO and BiTCN-BiGRU

Reliable sea level predictions are essential for ensuring the sustainability and ecological protection of coastal areas. An adaptive deep learning sea level height prediction hybrid model based on the improved dung beetle optimizer (OLSDBO), bidirectional temporal convolutional network (BiTCN), and bidirectional gated recurrent unit (BiGRU) is proposed in this paper. Initially, we optimize the BiTCN-BiGRU hyperparameters via OLSDBO. Sea level data are fed into the BiTCN, where bidirectional temporal convolutions with dilated causal layers and residual connections extract hidden information. Next, the extracted features are passed into the BiGRU to learn the dynamic changes in both directions, thereby capturing the temporal dependencies within the sequence. Finally, the optimal model prediction results are obtained. The model was evaluated via Australian tide gauge data and compared with nine relevant models. The experimental results show that the OLSDBO-BiTCN-BiGRU outperforms the comparison models, indicating its strong modeling capabilities. To address the randomness in neural network initialization, statistical comparisons were conducted with ten random seeds, confirming robustness. When applied to satellite altimetry data from the East China Sea, the model indicated a 3.28 ± 0.26 mm/a rise (1993–2023), corroborating the official bulletins. This study introduces a novel framework and practical pathway for regional sea level prediction, offering practical value for coastal management and climate adaptation strategies.