An Input-Conditioned Dynamic Hierarchical Spatiotemporal Graph Convolutional Network for Traffic Flow Prediction

Abstract

Traffic network partitioning studies show that road networks can be divided into spatially connected subregions with relatively homogeneous traffic states, whose composition may evolve with congestion. However, many hierarchical traffic forecasting models rely on offline or globally shared node-to-region relationships that remain fixed across input windows. We propose DH-STGCN, an input-conditioned dynamic hierarchical spatiotemporal graph convolutional network for traffic flow prediction. The model generates a window-specific soft node-to-region assignment from current spatiotemporal node representations, constructs corresponding regional representations and a regional graph, performs spatiotemporal learning at both sensor and regional scales, and feeds regional information back to the sensor scale. Aggregation consistency and region balancing are used as auxiliary structural regularizers. Experiments on PeMSD3, PeMSD4, PeMSD7, and PeMSD8 show that DH-STGCN reduces average MAE by 3.5–8.2% relative to STGCN and by 2.3–8.2% relative to HGCN. Relative to the strongest competing models, performance is dataset-dependent: Graph WaveNet achieves lower mean MAE on PeMSD3, PeMSD4, and PeMSD7, whereas DH-STGCN performs better on PeMSD8. Controlled hierarchy comparisons favor the window-conditioned assignment over fixed-uniform, static-hard, globally shared, and alternative differentiable assignments. Diagnostic analyses further show that the learned memberships are diffuse yet measurably nonuniform, change across consecutive input windows, and produce distinguishable regional representations. Overall, DH-STGCN provides a flexible input-conditioned hierarchical representation for multistep traffic flow prediction.