A vehicle trajectory prediction method based on variational automatic coder was proposed to address a series of problems such as inadequate modeling of the interaction between vehicles and surrounding vehicles and road geometry in trajectory prediction and incomplete multimodal modeling of vehicle trajectories. Firstly the semantic features of trajectory data and lane information were extracted from the original data through long and short term memory networks. Secondly a multi-headed attention mechanism was introduced and two separate attention modules were used to establish the vehicle-vehicle interaction model and vehicle-road interaction model respectively which can better reflect the interaction effects of surrounding vehicles and road geometry on vehicle trajectories and obtain rich scene context information. Next the vehicle trajectory was modelled multimodally by using the variational automatic coder to capture the random nature of trajectory prediction to generate a reasonable future trajectory distribution. Finally the sampling was repeated several times from the distribution to generate multiple possible future trajectories. By building the experimental platform and testing with the Argoverse naturalistic driving dataset the improved prediction method yielded values of 1. 03 and 1. 51 under the average displacement error and final displacement error indicators respectively and the prediction accuracy has been improved by 45% 46% and 32% compared with the other three prediction methods. The experimental results showed that the prediction method can effectively solve the problems of inadequate modeling of the interaction between vehicles and surrounding vehicles and road geometry and incomplete multi-modal modeling of vehicle trajectories. The prediction accuracy has been improved and the overall prediction performance is good.