In complex multi-vehicle dynamic interaction scenarios, intelligent vehicles must accurately perceive current risks to ensure driving safety. Existing vehicle trajectory prediction methods fail largely to capture the dynamic evolution of vehicle interaction risks and lack considerations of prediction self-consistency. This study proposes a continuous vehicle trajectory prediction method that explicitly accounts for dynamic interaction risks. First, vehicle states and interaction relationships are encoded as nodes and edges, respectively, to construct a graph structure. This graph sequence is then fed into an interaction-risk-aware spatio-temporal encoder, where the risk field is employed to represent vehicle interaction risk. A risk-guided graph attention mechanism, combined with the gated recurrent units, are used to extract spatiotemporal interaction risk features. Subsequently, the encoded information containing interaction risk features is passed to a self-consistent prediction attention mechanism, which leverages historical prediction information to refine the current embedded features. Finally, the adjusted features are input into a risk-adaptive continuous decoder to generate multiple plausible trajectories along with their distribution probabilities, accounting for driving uncertainty. Experiments on the HighD dataset show that the proposed method achieves a root mean square error of 0.85 m over a 5-second prediction horizon, with a 16% improvement in accuracy compared to the best baseline model. The results demonstrate that incorporating dynamic interaction risk enhances short-term vehicle trajectory prediction accuracy, while dynamically revisiting historical predictions and capturing dependencies between past and current predictions improves long-term vehicle trajectory prediction performance.