TCTT: Twin Convolution and Twin Transformers for Road Traffic Prediction

Road traffic prediction has emerged as a pillar of intelligent transportation systems, serving as a vital tool for mitigating congestion and enhancing the overall efficiency and fluidity of networks. Accurately forecasting traffic flow is inherently challenging due to the unpredictable nature of sudden local disruptions and the influence of large-scale global dynamics. The complex interplay between these factors makes the task exceptionally difficult, especially considering the highly dynamic and nonlinear behavior of traffic patterns. Nonetheless, numerous existing methods conflate these two layers of information, thereby limiting their effectiveness in representing the complex dynamics of traffic flow. In this study, we introduce a novel block-based framework designed to explicitly disentangle local and global dependencies across both spatial and temporal dimensions. Each block combines twin convolution for extracting multiscale spatiotemporal features and twin transformers for modeling complex hierarchical patterns. The integration of these components enhances both the accuracy and robustness of traffic prediction. The effectiveness of the proposed framework is validated on the los-loop and SZ-taxi datasets, demonstrating superior performance, particularly in long-term forecasting scenarios. The proposed approach yields RMSE reductions of 2.36%–41.57% on the SZ-taxi dataset and 2.45%–49.32% on the los-loop dataset, demonstrating superior performance over baseline models such as ARIMA, SVR, ASTGCN, T-GCN, and A3T-GCN.