A temporal causal deep learning for real-time traffic risk prediction on freeway merging areas

Conflict risk identification based on aggregate safety indicators often has low spatiotemporal resolution and provides limited support for proactive interventions. Additionally, deep learning models used in this context lack causal interpretability. To address these limitations, this study proposes a temporal causal deep learning framework for risk prediction in freeway merging areas. High-resolution vehicle trajectories collected by unmanned aerial vehicles are used to construct multivariate time series features that describe ego-vehicle behavior, car-following interactions with the lead vehicle, and multi-directional interactions with surrounding vehicles. A driving conflict risk index (RSI) is defined as the combination of Minimum Time to Collision (MTTC) and a crash impact measure to quantify both the likelihood and potential severity of longitudinal rear-end conflicts. Building on the iTransformer architecture, this study develops a TC-iTransformer model that integrates multi-scale temporal causal convolutions and multi-head causal attention to model lag-dependent relationships. The model further incorporates Regression Relevance Propagation (RRP) to infer cross-lag causal structures and characterize temporal interaction patterns among variables. Experiments on trajectory data from two freeway merging sites show that TC-iTransformer achieves the highest prediction accuracy, with an MAE of 0.602 and an RMSE of 0.910. Compared with a baseline model without the proposed causal modules, MAE and RMSE decrease by 17.3% and 14.1%, respectively. The interpretation results quantify key chain-like pathways and lag structures underlying RSI evolution, providing a time-resolved view of conflict-risk propagation. These findings support the identification of risky driving behaviors and the development of proactive safety management strategies in freeway merging areas.