A dual-channel ultra-short-term ship motion prediction model using temporal convolutional network, frequency-enhanced channel attention, and informer

Due to the complexity of the marine environment, ultra-short-term ship motion prediction has always been challenging. This study proposes a dual-channel network architecture, TF-Informer, which integrates temporal convolutional network (TCN), frequency-enhanced channel attention mechanism (FECAM), and Informer to improve prediction accuracy, abbreviated as TF-Informer. TCN handles short-term dependencies and extracts local features, while FECAM introduces frequency information to enhance the model's ability to capture frequency signals during feature extraction. Informer, with its self-attention mechanism, captures long-term dependencies by extracting global information. The parallel integration of these components enables the TF-Informer model to address complex nonlinear and non-stationary characteristics, capturing both short-term fluctuations and long-term trends, thus significantly improving prediction accuracy. This study conducted multiple tests using experimental ship model motion data, including multistep prediction, coupled data prediction, and generalization studies. The results indicate that, compared to the Informer model, the TF-Informer model reduces the mean squared error (MSE) by up to 29.53% in multistep prediction tests for accuracy. In complex and variable mixed sea conditions, the R2 value for single-step prediction approaches 1, and multistep predictions can also maintain high computational accuracy. In terms of generalization, the MSE of the forecast is reduced by up to 66.68%. Compared to the Informer model, the proposed TF-Informer model demonstrates significant improvements in prediction accuracy and generalization ability, offering a distinct advantage in ultra-short-term ship motion prediction in complex marine environments.