Coupled macro-micro modeling method for CAV lane-changing decision-making in V2X-enabled environments
TL;DR
A novel hierarchical decision-making and planning framework is presented, underpinned by a Spatiotemporal Attention Mechanism that adaptively weights and integrates macro-scale traffic flow information with micro-scale vehicle kinematics, offering a transparent and robust solution for intelligent vehicle decision-making in collaborative vehicle-road-cloud ecosystems.
Abstract
The operational efficacy of Connected and Automated Vehicles (CAVs) in mixed traffic flows critically depends on advanced decision-making systems capable of synthesizing high-dimensional environmental states. To address the limitations of existing methods in dynamically fusing lane-level traffic dynamics with vehicle-level interaction features, we present a novel hierarchical decision-making and planning framework. The framework is underpinned by a Spatiotemporal Attention Mechanism (STAM) that adaptively weights and integrates macro-scale traffic flow information with micro-scale vehicle kinematics. We trained a Deep Deterministic Policy Gradient (DDPG) agent as a high-level planner to generate cooperative lane-changing strategies, using multi-objective reward optimization. In a simulated highway scenario with 190 human-driven vehicles and 10 CAVs under varying penetration rates, results demonstrate that the proposed system effectively enhances mixed-traffic performance: under low CAV penetration (5%), ineffective lane-change maneuvers are reduced by up to 85.7%, while at medium-to-high penetration rates (30%–50%), cooperative interactions increase by a factor of up to 6.2. Across multiple independent runs, the hazardous TTC proportion remains within a narrow range, indicating a stable safety margin. The STAM also provides interpretable insights into what the system focuses on, offering a transparent and robust solution for intelligent vehicle decision-making in collaborative vehicle-road-cloud ecosystems.