Establishing an accurate dynamic model is crucial for the maneuvering and control of Autonomous Underwater Vehicles (AUVs). However, due to the highly nonlinear and strongly coupled hydrodynamic characteristics, it is a challenging issue to build a precise dynamic model. This paper proposes a Transformer-based method for AUV model identification that applies multi-head attention mechanisms independently to the various state variables and diverse temporal states of the AUV, hence extracting features from the input vehicle state data. A sliding window is utilized to continuously generate predictive data for assessing the effectiveness of the model identification. The performance of this method is verified in the Remus 100 underwater vehicle simulation system. The performance of this method is verified in the Remus 100 underwater vehicle simulation system.