Abstract
The construction of typical driving cycles is essential for evaluating vehicle energy consumption, yet existing methods struggle to balance search coverage, computational efficiency, and disturbance control. This paper proposes a novel method based on the Markov Chain Resampling Reachable Domain with Reinforcement Learning (RL-MC-RRD). Reinforcement learning jointly determines the sub-segments to be refined and their corresponding bridging-path search budgets, followed by state lookup table-based resampling to further correct the residual statistical discrepancies. Validated using one million seconds of real driving data from Qingdao, the final driving cycle constructed by RL-MC-RRD achieved an average deviation of only 0.206% across 16 characteristic parameters. Compared to traditional MC and MC-GA methods, it improves construction accuracy by 50-fold and 8-fold. In addition, the RL-MC-RRD-generated driving cycle exhibited deviations of 1.309% in electric-energy consumption and 3.075% in fuel consumption relative to the real-world driving cycle and offers an effective approach to accurate driving cycle construction, vehicle energy management, and performance evaluation.