Abstract: In the field of ocean science, tidal level prediction is an important research topic with wide application value. With the development of machine learning technology, machine learning methods have gradually become a new approach to tidal level prediction. To further explore the application of machine learning methods in long-term tidal level prediction, this paper constructs a machine learning model based on the Bidirectional Long Short-Term Memory (BiLSTM) network architecture. The hourly tidal level data from a total of 8 stations in the Changjiang Estuary region were input into the model for training, and a convolutional neural network layer was used to extract the periodic information of historical tidal levels over a period of 14 days, resulting in tidal level prediction models for each station. The experimental results show that in the 120-hour prediction task, the average Root Mean Square Error and Mean Absolute Error of the model are both below 20 cm and 17 cm, respectively, with a correlation coefficient higher than 0.98. The results demonstrate that the improved neural network model is effective in long-term tidal level prediction, significantly improving prediction accuracy and effectively mitigating the deficiencies of existing models in extracting tidal periodic features and addressing cumulative errors in long-term predictions.

Key words: Tidal Level Prediction; Machine Learning; Neural Network Model; BiLSTM; CNN