Comparing optimal and adaptive EV charging in smart cities: MILP vs. reinforcement learning

Abstract

The coordinated scheduling of electric vehicle (EV) charging is a critical challenge for smart cities, particularly in high-density infrastructure such as Mobility Hubs (MHs). This paper evaluates and compares two prominent approaches to the EV Charging Scheduling Problem (CSP): Mixed-Integer Linear Programming (MILP) and Reinforcement Learning (RL). We formulate a shared problem framework and apply both strategies under two structured scenarios: a small-scale deterministic benchmark and a medium-scale, realistic deployment with higher heterogeneity. Results show that MILP achieves optimal cost and state of charge SoC compliance in tractable cases but struggles with scalability. RL, based on Proximal Policy Optimization (PPO), achieves near-optimal performance while scaling to 100 EVs with minimal computation time. Despite occasional SoC deviations, the RL agent exhibits robust and adaptive behavior under dynamic conditions. This study offers actionable insights for selecting and deploying EV scheduling strategies in real-world urban environments.

This work was partially supported by the Spanish Government under this research project funded by MCIN/AEI/10.13039/501100011033: DISCOVERY PID2023-148716OB-C32. Also by the project MultiMO TSI100123-2024-0060; and by the predoctoral scholarship associated with the ”Generaci´on de Conocimiento” Projects, Call 2022, PRE2021-099830. Also, by the Generalitat de Catalunya AGAUR grant ”2021 SGR 01413”.

Peer Reviewed

Postprint (published version)