Abstract:

Balancing charging requirements with ride requests is essential for maximizing the operational revenue of autonomous vehicle (AV) fleets. To enhance service efficiency, this study proposes a Data-Driven Autonomous Scheduling (D-DAS) method for AV charging station leasing and nighttime relocation. By restricting charging to non-operational nighttime hours, the strategy maximizes daytime service availability and mitigates charging-related downtime. We develop a two-stage stochastic programming model that synergistically optimizes platform leasing decisions and fleet relocation strategies while accounting for future demand uncertainty. To address computational challenges, a Multilayer Perceptron (MLP) is employed as a surrogate model to map the complex nonlinear relationship between initial fleet distribution and operational revenue; this is then embedded into the lower-level model via mixed-integer linearization. Furthermore, Wasserstein distance-based scenario reduction transforms the stochastic model into a large-scale deterministic mixed-integer programming (MIP) problem, which is solved using an Integer L-shaped algorithm. Case studies using Manhattan taxi data demonstrate that D-DAS increases operational revenue by 15.1% and 19.8% compared to two benchmarks. Regarding investment, D-DAS saves approximately $300,000 and $1.8 million over heuristic and greedy methods, respectively. Under uncertainty, it yields an additional $6,500/day compared to deterministic models. These results prove that the proposed method effectively reduces costs, boosts profits, and ensures robustness, providing a theoretical and practical framework for platform infrastructure planning.

Key words: autonomous vehicle fleet, charging station leasing, nighttime relocation, two-stage stochastic programming