Significant Wave Height (SWH) exhibits complex nonlinear dynamic properties,posing significant challenges to its accurate prediction.Time-frequency decomposition is an effective approach to deal with such nonlinearities.However,existing methods fail to account for the distinct time-frequency characteristics of the SWH's decomposed components.Here,we employ permutation entropy to categorize SWH components,which are obtained via Ensemble Empirical Mode Decomposition (EEMD),into high- and low-frequency groups,and construct an optimized Long Short-Term Memory-Temporal Convolutional Network (LSTM-TCN) based on their respective characteristics,forming a dual-channel temporal feature extraction module.Furthermore,since different component predictions contribute unequally to the final SWH prediction result,the Bayesian Model Averaging (BMA) is introduced to assign adaptive weights.Finally,this paper proposes a dual-channel hybrid prediction model for SWH leveraging optimized deep learning.Experimental results show that,compared to state-of-the-art models,the proposed model achieves significant reductions in RMSE,MAE and MAPE across 1-,3-,6-,and 12-hour SWH predictions,with enhanced accuracy and stability.