Patch-Based Transformer–Graph Framework (PTSTG) for Traffic Forecasting in Transportation Systems

Accurate traffic forecasting underpins intelligent transportation systems. We present PTSTG, a compact spatio-temporal forecaster that couples a patch-based Transformer encoder with a data-driven adaptive adjacency and lightweight node graph blocks. The temporal module tokenizes multivariate series into fixed-length patches to capture short- and long-range patterns in a single pass, while the graph module refines node embeddings via learned inter-node aggregation. A horizon-specific head emits all steps simultaneously. On standard benchmarks (METR-LA, PEMS-BAY) and the LargeST (SD) split with horizons $\{3,6,12\}\to \{15,30,60\}$ minutes, PTSTG delivers competitive point-estimate results relative to recent temporal graph models. On METR-LA/PEMS-BAY, it remains close to strong baselines (e.g., DCRNN) without surpassing them; on LargeST, it attains favorable average RMSE/MAE while trailing the strongest hybrids on some horizons. The design preserves a compact footprint and single-pass, multi-horizon inference, and offers clear capacity-driven headroom without architectural changes.