Abstract
Road traffic crashes remain a significant concern for public safety and transport systems, and addressing their adverse effects forms a foundation for safety planning and policy development. This study presents a hierarchical hybrid framework that combines signal decomposition techniques, including Variational Mode Decomposition (VMD) and Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), with deep learning models: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Temporal Convolutional Network (TCN), and WaveNet. The framework uses daily vehicle–vehicle crash data from Yinzhou District, Ningbo City. Among all configurations, the VMD-GRU model produced the best results, with MAE = 2.960, RMSE = 3.750, and R(2) = 0.984, which reflects its ability to capture complex temporal structures. In contrast, the CEEMDAN-TCN model showed the weakest performance, with MAE = 14.559, RMSE = 19.481, and R(2) = 0.609. Furthermore, the Wilcoxon signed-rank test confirmed that the performance of VMD-GRU differs significantly from all other models at the 5% significance level. Residual analysis indicates that VMD-GRU maintains low prediction errors and aligns more closely with actual vehicle–vehicle crash values over time. This framework provides traffic authorities with a tool to identify shifts in crash patterns, make timely policy decisions, and allocate safety resources with greater precision.