Scalable ride-matching and dispatching model for shared autonomous electric vehicles with real-time demand prediction

Shared Autonomous Electric Vehicles (SAEVs) hold significant potential for reducing travel costs and alleviating traffic congestion. However, existing SAEV dispatch algorithms often employ shortsighted strategies and suffer from high computational complexity. This paper proposes a hierarchical framework for dynamic ride-matching and fleet dispatching that unifies demand forecasting and real-time decision-making. At the lower level, a high-precision passenger demand prediction model is constructed using an attention-based spatial–temporal graph convolutional network, leveraging hexagonal grid partitioning and historical travel data to generate fine-grained, time-sensitive forecasts. Building on these predictive insights, the upper-level decision module models the ride-matching task as a Markov decision process and synergizes this information within a multi-agent reinforcement learning framework, solved using an improved Double Deep Q-Network (Double DQN) algorithm. Experiments using real-world travel data from Chengdu in a multi-agent simulation environment demonstrate the effectiveness of the integrated framework. Results show that the demand prediction model achieves a root mean square error of 1.046, and the ride-matching decision method completes 93.76% of travel orders with only 76.6% of the SAEV fleet, significantly improving vehicle utilization. Additionally, when enabling both ride-sharing and demand-aware dispatching, as opposed to operating without these functionalities, the model achieves a 15% higher order completion rate. These results underscore the framework’s ability to couple high-fidelity forecasting with scalable decision-making, offering a robust and adaptive solution for SAEV-based urban mobility systems.