Optimal charging for large-scale heterogeneous electric vehicles: A novel paradigm based on learning and backward clustering

As the number of electric vehicles (EVs) surges, how to optimally manage the diverse charging behaviors of substantial heterogeneous EVs presents a significant challenge, as random charging activities may cause catastrophic consequences to disrupt power systems. While centralized control methods have garnered attention for their potential to achieve global optimality, they are primarily suitable for small or medium-sized EV fleets due to their high computational demands. In this work, to harness the benefits of centralized control methods while remarkably enhancing computational efficiency, a novel backward clustering scheme is introduced. The scheme is trained using optimal scheduling datasets collected from substantial cases involving small-scale EVs to categorize heterogeneous EVs with similar charging patterns but distinct states during dynamic charging. Subsequently, an online model predictive control is utilized to effectively manage distinct groups of EVs identified by the clustering method, rather than individual EVs. This control paradigm notably reduces the control dimension, alleviating the computational burden of centralized control. Simulation results demonstrate that the proposed control paradigm shows satisfactory performance, achieving near-optimal performance with only a 3% increase in total charging costs even in the case where 1000 EVs randomly participate in energy and ancillary electricity markets. More importantly, the calculation time can be reduced by 92% when compared to the benchmark methods, making its real-time implementation feasible.