As autonomous driving technology advances to higher levels, the demands for vehicle trajectory prediction in terms of accuracy, rationality, and safety have escalated significantly. This study focuses on optimizing traditional goal-oriented prediction methods. The proposed model incorporates high-precision map data and long-term temporal dependencies captured through the Transformer network. Leveraging these dependencies, the model can generate multiple potential target regions and meticulously assess the confidence level of each target area. Additionally, the algorithm takes into account the physical characteristics of vehicle dynamics equations and traffic rule constraints. By appropriately tuning parameters and weight distribution, it integrates these characteristics with a residual denoising diffusion model. The integrated model, through comparative experiments on the NGSIM dataset, achieved respective improvements of 4%, 14%, 3%, and 8% over the baseline model in \(minADE_1\) , \(minADE_5\) , \(minFDE_1\) , and \(minFDE_5\) . These results demonstrate that the optimized method enables the generated diverse vehicle trajectories to closely match the rational distribution of real-world traffic scenarios, effectively avoiding the generation of unreasonable trajectories.