Significant Wave Height (SWH) has complex nonlinear dynamic properties, which makes its accurate prediction a major challenge. Time-frequency decomposition technology is an effective means to deal with complex nonlinearities. However, existing methods do not consider the different time-frequency characteristics of the decomposed components of SWH. Therefore, this paper uses permutation entropy to divide the SWH components obtained after integrated empirical mode decomposition (EEMD) into high and low frequency categories, and constructs an optimized long short-term memory network - temporal convolutional network LSTM-TCN) based on their respective characteristics. Dual-channel temporal feature extraction module. Considering that different component prediction values have different impacts on the final SWH prediction result, the Bayesian average method (BMA) is introduced for weight allocation. Finally, this paper proposes a SWH dual-channel hybrid prediction model based on optimized deep learning. Experimental results show that compared with the existing state-of-the-art models, the evaluation indicators RMSE, MAE and MAPE of this model are significantly reduced in 1, 3, 6, and 12-hour SWH predictions, and it has better accuracy and stability.