Event-Driven Topology Reconfiguration and Penetration-Aware Graph Attention for Mixed-Autonomy Traffic Forecasting

Mixed-autonomy traffic with autonomous vehicles (AVs) and human-driven vehicles (HDVs) presents forecasting challenges because agent interactions and event-affected spatial dependencies vary over time. We propose Heterogeneous Adaptive Dynamic Spatiotemporal forecasting (HADS), a graph-attention framework that combines event-driven local topology reconfiguration, penetration-aware dynamic time warping (DTW) attention, and task-level temporal fusion for multi-horizon traffic-flow prediction. The forecasting target is traffic flow. The evaluation uses a semi-synthetic, penetration-controlled benchmark built from field-observed traffic-flow targets and 912 labeled anomalous events on a Beijing pilot-zone network (534 nodes and 3180 directed edges; April–July 2023), paired with SUMO-generated AV features at 20%, 40%, and 60% penetration. Under the reported single-seed runs, HADS obtains lower 15 min MAPE than AGCRN at 60% penetration, decreasing the point estimate from 2.92% to 2.61% under regular conditions and from 3.32% to 2.98% under anomalous conditions. Results across the reported 15 and 30 min settings indicate that event-conditioned topology and lag-aware heterogeneous attention can improve traffic-flow forecasting on this hybrid real–simulation benchmark. Results are from single-seed runs and should be read as preliminary point-estimate evidence rather than statistically demonstrated improvements; the semi-synthetic evaluation shows potential for pilot-zone applications rather than confirming real-world deployment performance.