An LLM-guided multi-objective optimization framework for cluster-aware electric vehicle charging management toward sustainable grid operation

The rapid growth of electric vehicles (EVs) is increasing stress on power distribution networks by amplifying peak demand and reducing charging-station utilization under uncoordinated charging. This study proposes a Large Language Model-guided Hyper-Heuristic (LLM-HH) for adaptive EV charging coordination. The framework uses a lightweight LLM-based meta-controller that monitors compact optimization feedback signals, including hypervolume (HV) improvement, population diversity, and stagnation, to orchestrate a portfolio of multi-objective evolutionary algorithms (MOEAs) through constrained solver selection and bounded parameter updates. Under an hourly coordination model, the framework coordinates a portfolio of MOEAs to jointly optimize three competing objectives: reducing on-peak demand, improving charger utilization, and preserving a cluster-level energy-per-session proxy across three behaviorally distinct charging regimes: Quick Top-Up, Home-Based, and Public-Stay. Experiments on a real-world Korean EV charging dataset demonstrate that the optimization-feedback-driven LLM-HH consistently improves Pareto front quality, achieving statistically significant improvements in Pareto-front quality over strong MOEA baselines and adaptive HH comparators, reducing on-peak demand by up to 18%, and incurring negligible runtime overhead ( < 2 % ). The framework also supports practical deployment through auditable, guardrail-constrained control actions grounded in optimization feedback. These results demonstrate that LLM-guided HH can improve the robustness, efficiency, and scalability of coordinated EV charging for more sustainable grid operation.