Attention Graph Neural Controlled Differential Equations Approach for Traffic Flow Forecasting

With the development of intelligent transportation systems, traffic flow prediction on complex traffic networks has become increasingly important, particularly in urban management and traffic planning, where it holds significant practical value. Existing traffic flow prediction methods primarily rely on stacking discrete modules to enhance prediction accuracy. However, the discontinuous hidden state trajectories between discrete modules may result in increased numerical errors and insufficient precision in capturing dynamic changes. Neural controlled differential equations (NCDEs), which can continuously adjust hidden state trajectories based on control signals, effectively address these issues. For this, we propose an Attention Graph Neural Controlled Differential Equations model (AG-NCDE). Specifically, the model first employs an attention controlled differential equation to model continuous-time dependencies. Then, it integrates high-dimensional spatiotemporal features through continuous dynamic spatiotemporal embeddings to construct a dynamic spatiotemporal graph. Finally, the spatial features are propagated using a graph controlled differential equation. The proposed AG-NCDE model is evaluated through comparative experiments on four public datasets: PEMS03, PEMS04, PEMS07, and PEMS08. The results demonstrate that the model exceeds most baseline models across three different metrics in terms of overall prediction performance. Additionally, it exhibits superior results in handling missing data.