Definition and evaluation of model-free coordination of electrical vehicle charging with reinforcement learning

Demand response (DR) becomes critical to manage the charging load of a growing electric vehicle (EV) deployment. Initial DR studies mainly adopt model predictive control, but models are largely uncertain for the EV scenario (e.g., customer behavior). Model-free approaches, based on reinforcement learning (RL), are an attractive alternative. We propose a new Markov decision process (MDP) formulation in the RL framework, to jointly coordinate a set of charging stations. State-of-the-art algorithms either focus on a single EV, or control an aggregate of EVs in multiple steps (e.g., 1) make aggregate load decisions and 2) translate the aggregate decision to individual EVs). In contrast, our RL approach jointly controls the whole set of EVs at once. We contribute a new MDP formulation with a scalable state representation independent of the number of charging stations. Using a batch RL algorithm, fitted $Q$ -iteration, we learn an optimal charging policy. With simulations using real-world data, we: 1) differentiate settings in training the RL policy (e.g., the time span covered by training data); 2) compare its performance to an oracle all-knowing benchmark (providing an upper performance bound); 3) analyze performance fluctuations throughout a full year; and 4) demonstrate generalization capacity to larger sets of charging stations