A Hybrid Valley Filling and NSGA-III Metaheuristic for Day-Ahead City-Scale Electric Vehicle Charging Scheduling

Electric vehicle (EV) fleets are expanding rapidly and will place substantial demand on distribution grids. Day-ahead scheduling of city-scale EV charging constitutes a constrained multi-objective optimization problem that must balance peak load, load variation, and valley utilization simultaneously. This paper proposes a structured warm-start strategy that embeds a load-conservation valley-filling (LCVF) heuristic into the NSGA-III metaheuristic, seeding the entire initial population with grid-compliant, valley-filling schedules before the first generation runs. This search-space shaping approach restricts the evolutionary search to a feasible subspace defined by LCVF, enabling convergence that random initialization cannot achieve within the same computational budget. On four seasonal city-level instances derived from real electricity consumption data from Campo Grande, MS, Brazil ($N=50,336$ vehicles), VF–NSGA-III reduces peak load by $0.54$–$2.52\%$ (mean $1.31\%$) relative to standalone LCVF while requiring only $\approx 1.5\%$ of its runtime. The warm-start provides a structural advantage that population scaling alone cannot overcome: LCVF-initialized NSGA-III with $N_{\mathrm{pop}}=10$ achieves a hypervolume $35\%$ above the randomly initialized variant with $N_{\mathrm{pop}}=100$. A 32-day generalization study (June 2022–May 2023) confirms a mean peak-load reduction of $4.91\%$ over standalone LCVF and $4.93\%$ over randomly initialized NSGA-III across all seasons, demonstrating consistent performance over a full annual demand cycle.