<p>Vehicle trajectory prediction is a critical technology in autonomous driving, enhancing decision-making accuracy by forecasting the motion trajectories and behavioral intentions of surrounding agents through deep neural networks. Despite the advances in trajectory prediction models, significant challenges remain: current models inadequately capture the interactions between agents; the complexity of high-dimensional spatiotemporal data hinders feature learning; and trajectory prediction is inherently multimodal-identical historical trajectories can lead to varied future outcomes. To address these issues, this study proposes MSTD, a multimodal trajectory prediction method using multi-level spatiotemporal decoupling and spatiotemporal synchronous graph convolution. MSTD decouples different time series tasks across three levels-temporal, spatial, and spatiotemporal-allowing for the extraction of temporal features, spatial features, and critical spatiotemporal features, thereby enabling potential deep feature extraction from complex spatiotemporal data. To fully capture the interaction information among agents, a Spatial Attention Interaction Module is introduced, employing the multi-head attention mechanism from the Transformer to investigate interactions between vehicles at the same time step. Furthermore, multiple structurally consistent decoders are employed to generate predicted trajectories, with confidence levels accompanying each. The experimental section outlines the model’s parameter configurations and performance evaluations on datasets Argoverse. Results indicate that the minimum Average Displacement Error (minADE) of MSTD achieves 1.15, the minimum Final Displacement Error (minFDE) achieves 0.75, and the Miss Rate (MR) achieves 0.10, demonstrating a significant improvement in prediction accuracy over mainstream algorithms.</p>