In the Internet of Things era, coordinating electric vehicle (EV) charging schedules to balance user satisfaction, station profit, and grid stability constitutes a high‑dimensional multi‑objective challenge. Existing multi‑agent reinforcement learning (MARL) methods often falter due to the explosion of state spaces from massive sensing data and conflicting stakeholder interests. This paper introduces, for the first time, an LLM‑enhanced MARL framework that jointly optimizes the grid, EVs, and charging stations within a unified loop.
The core innovation lies in leveraging a Large Language Model (LLM) to overcome two critical bottlenecks: (1) interpretable feature selection—LLM semantically parses real‑time IoT states to extract physically meaningful features; (2) adaptive multi‑objective weighting—LLM uses semantic reasoning to dynamically assign weights to conflicting goals such as profit, user satisfaction, and grid load, eliminating cumbersome manual tuning.
Extensive experiments on a large‑scale urban charging network simulator compare the proposed approach against state‑of‑the‑art baselines. Results demonstrate superior market efficiency and a reduction of training time by over 70%. The solution is scalable and transparent, offering a novel pathway for sustainable IoT‑enabled urban charging infrastructure management.
Review