Abstract

The prediction of ocean wave heights and periods is crucial for reducing maritime accidents. The application of neural networks to predictive modeling eliminates the need for complex parameter schemes that describe specific physical processes, thereby reducing computational costs. This study presents two single and three fusion models based on neural-network architectures: Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), LSTM with an attention mechanism (LSTM+Attention), LSTM fused with a CNN (CNN+LSTM), and LSTM with Ensemble Empirical Mode Decomposition (EEMD+LSTM). The models were compared on multiple prediction lengths based on multivariate datasets from two oceanographic stations southeast of Fujian Province. All models performed well for prediction lengths of less than 24 h, with the coefficients of determination for wave height and period prediction higher than 0.7 and 0.6, respectively. For prediction lengths longer than 24 h, the predictive capability of EEMD+LSTM was superior to that of the other models. For a 48 h prediction length, the predictive results at each time step were compared with the true values across four typhoons. At the 24th and 48th time steps, phase lags were observed in the results from the CNN and LSTM. In contrast, EEMD+LSTM better captured the trends of the true values.

IPC Classification

Keywords