A data-driven prediction model for vehicle travel risks in tunnels

To enhance risk identification and prediction accuracy in tunnel vehicle operations, this study proposes a BO-GRU (Bayesian Optimization-Gated Recurrent Unit) model. The model is validated using vehicle trajectory data from Tongji University Road Trajectory Data Platform, achieving a risk identification accuracy of 99.88% and a potential risk prediction accuracy of 98.89%. Under optimal input conditions, trajectory prediction yields root mean square errors of 9.3314 in the longitudinal direction and 0.0470 in the lateral direction. Comparative experiments show that BO-GRU performs on par with parameter-optimized LSTM networks, while requiring fewer parameters and offering better computational efficiency. These results highlight the necessity of hyper-parameter optimization for data-driven models. The proposed framework supports real-time risk assessment for individual vehicles and predicts future risk trends across tunnel zones, enabling dynamic monitoring and early warning. The findings provide methodological insights for intelligent tunnel management and underscore the potential of integrating optimization techniques with recurrent neural networks to improve prediction accuracy and operational safety in complex traffic environments.