NATTrek: A network-aware trajectory gamified visual framework for cooperative EV fleet charging under feeder constraints

초록

Coordinated electric-vehicle (EV) fleet charging is increasingly constrained by shared feeder capacity, heterogeneous departure requirements and uncertain electricity prices. This study proposes NATTrek, a Network-Aware Trajectory gamified visual learning framework for cooperative EV fleet charging under feeder constraints. NATTrek reformulates fleet control as a structured visual decision environment in which day-ahead and real-time prices, forecast uncertainty, feeder utilization, EV connection windows, state-of-charge trajectories, target requirements, past actions, mission status, and dynamic priority scores are encoded in a compact gameboard. The visual observation is combined with numerical features, feasibility masks, and cooperative rewards to support imitation learning and multi-agent reinforcement learning while preserving operational constraints. DAgger, QMIX, and MAPPO are evaluated within the NATTrek interface against Immediate charging, Priority-Greedy coordination, and fleet oracle. Experiments use paired held-out episodes based on CAISO price trajectories, fleet sizes of 5, 10 and 20, and moderate and tight feeder-capacity regimes. Results show that hybrid learned policies improve departure reliability relative to their raw learned counterparts, but this improvement should be interpreted as the effect of a hybrid learned controller rather than a purely neural policy. Among the hybrid policies, QMIX-Hybrid gives the lowest SoC gap in all fleet-size and feeder-regime combinations. At N=10, QMIX-Hybrid achieves a SoC gap of 1.16% under the moderate feeder regime and 3.31% under the tight feeder regime. Its paired cost gain over Immediate charging is 25.0% in the moderate regime and 72.5% in the tight regime. Ablation results further show that the visual gameboard contributes mainly to cost performance and interpretability, while numerical features, local feasibility masking, feeder admission, and the Hybrid fallback are the main sources of reliability. These findings indicate that visual, priority-aware, and feasibility-masked learning can serve as an aggregator-side decision-support layer for managed charging sites, where real-time charger status, feeder headroom, electricity prices, and EV departure requirements are already available to an energy-management system.

키워드