Design of fair and interpretable electric vehicle charging policies through genetic programming

Controlled charging of a group of electric vehicles (EVs) subject to power limits, such as those imposed by transformer capacity constraints, may result in inefficient and unfair energy distribution among EVs. The present work aims to learn fair and efficient charging policies on historical data using multi-objective genetic programming (GP). A formula, or model, is evolved through GP that takes features of connected EVs as input variables and computes scores to guide the distribution of available energy. Two variants of this approach are proposed and evaluated in simulation experiments, considering a residential charging scenario. In this setting, the dissatisfaction of EV users resulting from a certain charging control policy is quantified as the additional time they required for charging externally, compared to uncontrolled charging. The efficiency and fairness of a charging control policy are measured as the mean and maximum additional external charging time over all users. In the simulation experiments, the proposed approach is compared to several baseline methods, including different manually designed charging policies, such as equal distribution and first-come-first-served, as well as a recent approach for combining multiple fixed charging policies. The experimental results show that the proposed approach increases efficiency by at least 13 % and fairness by at least 12 %. An analysis of the automatically designed policies in terms of interpretability concludes that the best-performing policies are highly complex, containing more than 13 variables and more than 16 operators, on average. However, it is shown that it is possible to significantly reduce this complexity without substantial loss in the quality of the charging control, through appropriate control of the maximum GP tree size.