Abstract: Ship trajectory prediction has become increasingly important in marine traffic management, shipping safety, and related fields. Current methods for ship time-series prediction exhibit certain limitations when handling multi-feature data inputs in water traffic scenarios, as they fail to adequately capture the correlations among features or focus on the critical information within time-series data. To address these shortcomings and further improve the accuracy of ship trajectory prediction, this study proposes a method named GCAU, which integrates an improved Graph Convolutional Network（GCN）with a Recurrent Attention Unit（RAU）. First, Graph Convolutional Networks are employed to capture interdependencies between features, thereby enhancing the model’ s capability to extract feature correlations. Second, an attention gate is incorporated into the Recurrent Attention Unit（RAU）, enabling selective emphasis on time-level features. Finally, the study evaluates four different ship time-series prediction methods across three distinct scenarios. The results demonstrate that GCAU outperforms the other methods in all tested scenarios, achieving lower values in Mean Squared Error（MSE）, Root Mean Square Error（RMSE）, Mean Absolute Percentage Error（MAPE）, and Mean Absolute Error（MAE）. The proposed method can effectively enhance the accuracy and stability of ship trajectory prediction, thereby providing more reliable decision support for maritime traffic management and other related applications.