Highlights • A Spatiotemporal Transformer Network is proposed for simulating conflict risk propagation on freeways. • The proposed architecture can capture the dynamic spatiotemporal dependencies among conflict risk sequences. • Conflict risk propagation simulating based different SSMs are conducted and compared. • The PSD emerges superiority over the other SSMs for conflict risk propagation simulating. • The proposed approach outperforms the selected several state-of-the-art machine-learning models.

Abstract On freeways, sudden deceleration or lane-changing by vehicles can trigger conflict risk that propagates backward in a specific pattern. Simulating this pattern of conflict risk propagation can not only help prevent crashes but is also vital for the deployment of advanced vehicle technologies. However, conflict risk propagation simulation (CRPS) on freeways is challenging due to the nuanced nature of the pattern, intricate spatio-temporal interdependencies among sequences and the high-resolution requirements. In this work, we introduce a conflict risk index to delineate potential conflict risk by aggregating various surrogate safety measures (SSMs) over time and space, and then propose a Spatio-Temporal Transformer Network (STTN) to simulate its propagation patterns. Multi-head attention mechanism and stacking layers enable the transformer to learn dynamic and hierarchical features in conflict risk sequences globally and locally. Two components, spatial and temporal learning transformers, are innovatively incorporated to extract and fuse these features, culminating in a fine-grained conflict risk inference. Comprehensive tests in real-world datasets verified the effectiveness of the STTN. Specifically, we employ three widely-recognized SSMs: Modified Time-To-Collision (MTTC), Proportion of Stopping Distance (PSD), and Deceleration Rate to Avoid a Collision (DRAC). These SSMs, gleaned from vehicle trajectories, are employed to delineate the conflict risk. Then, we conduct three comparative simulation tasks: MTTC-based model, PSD-based model, and DRAC-based model. Experimental results show that the PSD-based model exhibits a robust performance on all tasks, and is minimally affected by the durations of the simulation time, while the DRAC-based model more distinctly delineates the spatio-temporal conflict risk heterogeneity. Furthermore, we benchmark the STTN against three common state-of-the-art machine learning models across all simulation tasks. Results reveal that the STTN consistently surpassed these benchmark models (LSTM, CNN and ConvLSTM), suggesting the potential of the attention mechanism on the CRPS tasks. Our investigation offers crucial insights beneficial for traffic safety warning, advanced freeway management systems, and driver assistance systems, among others.