NAMD-Net: A Dual-Branch Model for Significant Wave Height Prediction Based on Noise-Assisted Multivariate Empirical Mode Decomposition

Accurate prediction of significant wave height (SWH) is essential for ocean engineering safety, maritime transportation, and wave energy utilization. To address the nonlinear and non-stationary characteristics of SWH time series, this paper proposes a hybrid forecasting model that integrates signal decomposition, mode grouping and reconstruction, and deep learning. First, the model employs noise-assisted multivariate empirical mode decomposition (NA-MEMD) to adaptively decompose the multivariate time series into intrinsic mode functions (IMFs) to alleviate the problem of mode mixing. Subsequently, permutation entropy is employed to assess the complexity of the mode components and guide their grouping into two branches. Finally, WaveNet and long short-term memory (LSTM) networks are used to model and forecast the relatively high-complexity (fast-varying) and low-complexity (slow-varying) components, respectively, and the final forecast is generated by superimposing the predictions of all components. The proposed model was validated using measured data from two buoy stations, and evaluated by four metrics: root mean square error (RMSE), mean absolute error (MAE), symmetric mean absolute percentage error (sMAPE), and correlation coefficient (R). Experimental results show that the proposed model achieves lower RMSE, MAE, and sMAPE, and higher R compared with the baseline models across all evaluated forecast horizons and buoy sites.