In maritime navigation, ships operating in open waters encounter complex dynamics that result in irregular six-degree-of-freedom motion responses. The inherent unpredictability and randomness of these movements pose substantial uncertainty and risks to various maritime operations, including missile launches and helicopter takeoffs and landings. Therefore, the ability to accurately predict a ship’s motion attitudes within a short timeframe is of critical importance for enhancing safety in navigation and operational efficiency. This paper proposes an advanced methodology for predicting ship motion attitudes based on an enhanced Transformer model. The proposed framework optimizes the original Transformer architecture by integrating a convolutional layer into the encoder module, which facilitates more comprehensive feature extraction and enriches the representation process. Through the synergistic interaction between this convolutional layer and the multi-head attention mechanism, deeper correlations within time series data are effectively examined. Experimental results demonstrate that this improved Transformer-based approach for predicting ship motion attitudes achieves high accuracy levels, underscoring its practical relevance in engineering applications.