Deep Learning-Based Mixed-Integer Co-Optimization of Velocity Planning and Powertrain Control in Urban Environment

This paper proposes an online co-optimization strategy for velocity planning and powertrain control in autonomous vehicles (AVs), aiming to improve energy efficiency and travel time in urban driving. A hierarchical eco-driving framework is designed to incorporate both long-horizon traffic signal information and short-horizon traffic dynamics. In the upper layer, velocity planning is formulated as a mixed-integer optimal control problem (MIOCP) in the distance domain. Binary variables are introduced using the big-M method to model disjunctive constraints associated with signalized intersection crossing decisions. The lower layer ensures velocity eco-tracking and car-following safety in the time domain while co-optimizing gearshift, traction, and braking torque, also resulting in an MIOCP. To address the real-time computation challenge, a two-stage optimization method is introduced. First, a neural network (NN) is trained offline via supervised learning to predict feasible integer strategies based on MIOCP parameters. Then, the predicted integers are fixed online, and the resulting problem is solved as a nonlinear programming (NLP) problem. The effectiveness of the proposed method is thoroughly validated through extensive simulation studies under various urban scenarios, including multiple signalized intersections, varying road speed limits, and diverse preceding vehicle (PV) behaviors. The results demonstrate that the proposed approach enables safe and adaptive eco-driving, obtaining a 5.97% average improvement in energy efficiency over the intelligent driver model, while accelerating computation speed by 1–2 orders of magnitude compared to Bonmin and achieving millisecond-level computational time.