Research on Trajectory Prediction and Risk Control for Distributed-Drive Electric Vehicles Based on a Spatiotemporal Transformer and an Improved DRA Algorithm

Considering the strong nonlinearity and dynamic coupling characteristics of distributed-drive electric vehicle trajectories, as well as the slow convergence and susceptibility to local optima of conventional optimization algorithms when tuning deep temporal models, this paper proposes a trajectory prediction and risk identification method based on an improved Divine Religions Algorithm (IDRA)-optimized spatiotemporal Transformer. First, an adaptive computation-saving mechanism and a dynamic community interaction mechanism are introduced into DRA to reduce redundant fitness evaluations, maintain population diversity, and improve convergence performance. Second, a spatiotemporal Transformer integrating temporal self-attention, spatial interaction attention, and spatiotemporal feature fusion is constructed to capture vehicle motion characteristics and multi-vehicle interaction dependencies. IDRA is then employed to automatically optimize key Transformer hyperparameters. Experimental results on eight benchmark functions show that IDRA achieves approximately 1.5-3 times faster convergence and improves optimization accuracy by 1-15 orders of magnitude compared with the comparison algorithms. On the self-constructed Multi-Vehicle Interaction Trajectory Dataset for Distributed-Drive Electric Vehicles, the proposed method achieves an MAE of 0.03 m, RMSE of 0.04 m, ADE of 0.02 m, and FDE of 0.05 m, outperforming GAT, GNN, and CNN-LSTM across all evaluated trajectory prediction metrics. These results demonstrate that the proposed method can effectively improve trajectory prediction accuracy under complex multi-vehicle interaction scenarios and provide reliable trajectory information for early risk identification and subsequent active-safety decision-making.