In this research, an optimization method for wave intelligent prediction models based on frequency-band preprocessing is proposed. The real sea experiment is carried out to collect wave data by the wave rider buoy. Based on the measured wave data, the temporal correlation features for different wave frequency bands are analyzed in this paper. Then the optimization effects of the ANN-WP, LSTM-WP and TCN-WP models under different sea states are further explored. Results indicate that the frequency-band preprocessing method proposed in this paper can effectively enhance the accuracy of wave prediction models and exhibits excellent adaptability to different sea states and neural network models.
In real-sea conditions, the ship will suffer from six-degree-of-freedom (6-DOF) swaying motions due to the complex environmental factors when sailing, which has great influence on the safety of ship navigation and operations. As the main environmental disturbances, the ocean wave plays a crucial role in determining the operational state of the ship. Therefore, the deterministic prediction of the wave is fundamental to provide control tactics and ensure the safety of maritime operations.
The wave prediction has a development history. The early research mainly focused on the statistical characteristics prediction of waves, such as the commonly used WAM (Komen G J et al., 1994), SWAN (Tolman H L et al., 2002), and WAVEWATCH III (Booij N et al.., 1999). These prediction methods are mainly based on the energy distribution of ocean waves, which can provide long-term statistical characteristics of the wave surface in the large filed. But it is difficult to provide refined wave elevation at fixed points in space. With the rapid development of ocean engineering, the phase-resolved wave prediction method, which can predict the accurate wave surface change in the future periods, has gradually become the key research filed.
In the development of the phase-resolved wave prediction method, the hydrodynamic method has always been a research hotspot. It realizes the wave prediction through the decomposition and reconstruction of the sea surface by measuring the velocity distribution of the wave field (Huchet M et al., 2021) or the time distribution of the wave surface (Morris, E. L. et al., 1998). But the wave velocity distribution is difficult to directly observe in the real sea, the methods based on wave surface distribution is used more commonly. Belmont et al. (2009) pointed out that the wave data with obvious narrow-band distribution characteristics can be effectively decomposed by the Discrete Fourier Transform (DFT). On this basis, Halliday et al. (2011) successfully predicted the wave surface by superimposing the wavelet component after DFT at the specific position based on the linear wave theory. This prediction method based on DFT has achieved excellent prediction results in many studies (Halliday et al. 2005, Simanesew. 2013, Fisher et al. 2021). Schrodinger equation (Simanesew et al., 2017) and Zakharov equation (Stuhlmeier et al., 2021) are also applied for the wave prediction, and it has been proved that they can achieve good prediction results for long-crested waves. Hydrodynamic prediction methods are mainly using linear wave theory (Wijaya et al., 2015) and nonlinear wave theory (Wang et al., 2022) calculate the wave propagation process. The linear model has high computational efficiency, but it is difficult to obtain good prediction accuracy for real sea waves, which contains obvious nonlinear characteristics. For the nonlinear model, although it can achieve high prediction accuracy, it costs more computation resources. Therefore, the hydrodynamic prediction method has the problem that it is difficult to balance the computational efficiency and accuracy.