Bayesian optimization for battery electric vehicle charging station placement by agent-based demand simulation

Abstract

The optimal placement of electric vehicle charging infrastructure (EVCI) is critical for meeting the charging demand of battery electric vehicles (BEVs). However, the questions of how to coordinate the stochastic charging demand and placement of charging infrastructure in an economically effective way and how to solve such a problem on a large real-world scale remain unanswered. This study addresses these questions by developing an agent-based charging station placement model (ACPM) that incorporates demand uncertainty by simulating the behaviors of BEV users. A solution algorithm based on random embedding Bayesian optimization (REMBO) is proposed and shown to significantly improve computational efficiency and solve the ACPM on a large real-world scale. The model and algorithm are applied to a case study of the Atlanta metropolitan area. In the study, the algorithm finds an optimal placement using only 2% of the runtime of existing benchmark methods. This enables the use of the algorithm on a more complex, real-world scale. Results from the case study indicate that the optimal placement of EVCI uses level-2 and direct-current fast charging (DCFC) roughly equally and distributes these relatively evenly throughout the region. This optimal EVCI placement has significant economic benefits because it can improve the expected net present value (NPV) by 50% to 100% relative to common benchmark charging placement practices, such as placing the EVCI uniform randomly or weighted by the number of parked vehicles. Finally, a sensitivity analysis demonstrates that factors such as the size of the BEV market, charging preferences, and charging price are shown to have significant impacts on the optimal placement and profitability of an EVCI project.