A multi-target trajectory cooperative prediction model (RGMA) based on a dynamic risk map and multivariate attention mechanism was proposed aiming at the problems of insufficient accuracy and generalization ability in vehicle trajectory prediction in complex traffic scenarios. A dynamic risk graph that integrated multi-factor interaction features such as vehicle size, speed, acceleration, and heading angle was constructed, and the conflict risk between vehicles was quantified as the adjacency weight of the graph convolutional network, enhancing the physical interpretability of spatial interaction modeling. A multivariate attention Transformer module was designed to treat the time series of each variable as an independent token, capturing cross-variable dependency and long-term temporal feature in order to improve temporal modeling capability. The future trajectories of multiple vehicles were output through concatenating the spatiotemporal feature and a multilayer perceptron. Experiments on real-world dataset NGSIM and HighD show that RGMA outperforms existing mainstream methods in both short-term and long-term prediction. Ablation study verifies the effectiveness of each module and the robustness of the model.