Traffic-Flow Prediction Using T-AGCN Based on an Adaptive Search Graph Fusion Mechanism

Abstract

The rapid growth in traffic volumes has increased the demand for traffic-flow forecasting models with stronger prediction capability. Traditional methods that rely on local feature extraction and static spatial graph construction can no longer fully meet these requirements. To address the short- and long-term fluctuations in freeway traffic flow and the dynamic correlations among node flows, this study proposes ASGF-TAGCN, a traffic-flow forecasting model based on a component-specific multi-source graph fusion mechanism. The model integrates CEEMDAN-based multiscale decomposition, multi-source graph construction (physical topology, DTW-based semantic similarity, and node-adaptive learning), a learnable fusion mechanism for multi-source spatial modeling, Transformer-based temporal modeling, and IHPO hyperparameter optimization. Across three independent training runs at the 30 min forecasting horizon, ASGF-TAGCN achieved a mean MAE of 2.2982 ± 0.1181 and a mean RMSE of 3.1147 ± 0.1059, where the variability is reported as the sample standard deviation. In the fixed seed-42 run used for the baseline comparison, ASGF-TAGCN achieved an MAE of 2.219 and an RMSE of 3.063, reducing the two errors by 14.72% and 11.58%, respectively, relative to T-AGCN, the strongest baseline among the selected models. These results demonstrate that ASGF-TAGCN consistently reduces prediction errors and effectively captures complex spatiotemporal dependencies, offering a reliable solution for short-term freeway traffic-flow forecasting.