Disturbance-aware hybrid signal–vehicle control for service maintenance in mixed-autonomy urban networks

Mixed-autonomy urban networks may experience rapid service degradation when local capacity losses interact with demand surges and queue spillback. This study formulates disturbance-period traffic management as a service-maintenance problem and proposes Res-STC (Resilience-oriented Signal-Traffic Cooperative Control), a hybrid multi-agent reinforcement-learning framework that coordinates network-level signal decisions with local speed guidance for eligible connected and autonomous vehicles (CAVs). A cascaded graph-attention/Transformer encoder represents adjacent-intersection coupling and longer-range network dependencies, while an engineering-constrained execution layer enforces minimum and maximum green times, yellow/all-red transitions, and directional anti-starvation rules. Experiments were conducted in SUMO on a 4 × 4 Hangzhou benchmark network adapted from the network used in CoLight, comprising 16 signalized intersections, under 13 demand, disturbance, and CAV-penetration scenarios. Across 10 matched simulation seeds for a selected trained policy, Res-STC achieved the highest or near-highest mean operational performance in most scenarios, particularly in average speed, throughput, and completion rate. Relative to the signal-only ablation, the scenario-level mean change in average speed ranged from − 0.01% to 4.90%, and the corresponding mean change in throughput ranged from − 0.29% to 3.89%. These descriptive results indicate that network-level signal coordination is the primary source of improvement, whereas CAV speed guidance acts as a complementary, scenario-dependent local actuator rather than a substitute for signal control.