Learning-augmented hierarchical control for signal-aware safe eco-driving of connected autonomous vehicles

Connected autonomous vehicles (CAVs) can significantly enhance energy efficiency by adjusting speed profiles in response to signalized intersections using Vehicle-to-Infrastructure (V2I) information. However, real-time eco-driving remains challenging due to complex constraints, smoothness requirements, and limited onboard computational capacity. This paper proposes a learning-augmented hierarchical control (LAHC) framework for signal-aware safe eco-driving in multi-intersection urban scenarios. LAHC combines open-loop trajectory planning via model predictive control (MPC) and closed-loop real-time control using an input convex neural network (ICNN), trained to approximate short-term optimal control inputs under dynamic disturbances. To ensure safety and feasibility, a lightweight quadratic safety filter (QSF) is integrated to refine the ICNN’s outputs with minimal overhead. Compared with the baseline method used for supervision, LAHC achieves a 1.43 % reduction in energy consumption and 51.2 % lower average computation time on the EV model. Relative to a recent state-of-the-art (SOTA) approach, it improves inference speed by 76.69 % while maintaining or improving energy performance. In other test conditions, LAHC produces results close to baseline but with significantly reduced computation time. These results demonstrate the real-time feasibility, safety, and energy efficiency of LAHC, offering a practical solution for safe eco-driving in complex urban environments.