Development of a refined wave field forecast model based on LSTM and its application in the Qiongzhou Strait

This paper develops a refined wave field prediction model based on Long Short-Term Memory (LSTM) networks. Enhancements in the dataset and feature selection are implemented to improve predictive performance. The dataset is derived from a self-moored buoy in the Qiongzhou Strait. The ERA5 reanalysis dataset is used for validation of the buoy's measurements. To enhance the model’s prediction efficiency, additional models are built using Random Forest and XGBoost, where the feature importance for time steps t-3, t-6, and t-12, relative to the current time (t), is ranked for predicting average wave height. Comparative models based on Recurrent Neural Networks (RNN) and Gated Recurrent Units (GRU) are also constructed for wave height prediction. The evaluation criteria— Mean Square Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the Coefficient of Determination (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>)—confirm that the proposed model outperforms other comparison models, achieving the best results when predicting the next hour's average wave height using data from the previous three hours.