Large-scale traffic safety management using Temporal Fusion Transformers: Prediction, interpretation, and intervention analysis

As urban road networks become increasingly complex, managing safety is a critical challenge. Proactive safety management requires both accurate risk prediction and an understanding of contributing factors. This study introduces a Temporal Fusion Transformer (TFT)-based framework that integrates prediction, interpretation, and intervention analysis for large-scale safety improvement. The TFT predicts a risk index for each intersection and link, utilizing both static and dynamic variables. Its dynamic attention mechanism provides insights into how the effect of each variable shifts over time. Trained and validated on the high-resolution pNEUMA trajectory dataset from Athens, Greece, the model achieved strong predictive performance (MAE = 0.30, RMSE = 0.461). Temporal analysis of attention weights revealed that taxis and motorcycles consistently contribute the most to elevated safety risks, while medium vehicles were most influential during the early morning hours. These insights enable the development of targeted strategies, such as dynamic access restrictions or time-based charges for specific vehicle classes during high-risk periods. A sensitivity analysis further shows how modest changes in traffic composition can significantly improve safety, quantifying the benefits of proposed interventions. Overall, the framework offers transportation agencies a practical, data-driven tool that bridges the gap between advanced predictive modeling and real-world safety management strategies.