BGAR: A Dual-Channel Deep Learning Framework for Urban Expressway Traffic Accident Prediction

Abstract

Accurate accident prediction is crucial for proactive safety management on urban expressways. However, its practical efficacy is hindered by several complex challenges, including the heterogeneity of causal data, the need to model the full temporal evolution of risk, and the synergistic, non-linear interactions between variables. To address these challenges, this study proposes BGAR, a dual-channel deep learning framework. The framework features a dual-channel architecture to disentangle static and dynamic data streams, a bidirectional GRU to model the complete risk lifecycle, and a multi-head attention mechanism to weigh critical factor combinations. Validated on a real-world expressway dataset, BGAR demonstrates superior predictive accuracy, outperforming the strongest of 12 established baseline models by 3% in terms of $R^2$. More importantly, it provides a diagnostic tool that translates forecasts into actionable control strategies. By pinpointing risk drivers, the framework enables a fundamental shift from reactive response to precise, proactive safety management, thus bridging the gap between prediction and prevention. The predictive target is the short-term accident count for the monitored corridor, enabling operators to quantify imminent risk levels in addition to identifying their drivers.