Forecast-guided multi-task scheduling for autonomous electric ride-hailing fleets under charging and maintenance constraints

Autonomous electric ride-hailing fleets require real-time coordination among passenger dispatching, proactive repositioning, charging, and maintenance. This paper proposes a forecast-guided rolling framework that combines an attention-based spatial-temporal graph convolutional network (ASTGCN) with Enhanced Kuhn-Munkres (EKM) multi-task assignment. ASTGCN predicts grid-level demand, while a virtual-node (VN) mechanism converts forecasted regional shortages into bounded point-level relocation tasks at feasible points of interest (POIs). Passenger orders, VN-based repositioning, charging, and maintenance are then coordinated in a common vehicle-task graph after infeasible vehicle-task pairs are filtered according to battery, health, time-window, charging-capacity, and maintenance-capacity constraints. A residual-based search mechanism reduces the computational burden introduced by graph expansion. Experiments using real-world ride-hailing data show that the proposed framework achieves a system matching rate of 79.76%, which is 5.52 percentage points higher than the classical Kuhn-Munkres matching baseline. It reduces average passenger waiting time to 6.68 min, corresponding to a 14.41% reduction compared with the non-predictive scheduling variant. It also improves spatial supply-demand balance, reducing the normalized demand-supply difference to 0.128, and completes each rolling scheduling decision within 1.42 s on average. The incremental ablation study further shows that virtual-node-based forecast conversion mainly improves service quality and spatial balance, while the residual-based search mechanism reduces computation time from 22.45 s to 1.42 s without observed degradation. These results indicate that the main value of the proposed framework lies in converting demand forecasts into executable multi-task assignments for real-time autonomous electric fleet operation.