Abstract

With the increasing penetration of electric vehicles (EVs), the orderly charging-discharging (C-D) strategy of EVs can effectively alleviate the impact of large-scale grid-connected EVs. However, the schedule of the C-D system of EV clusters encounters the curse of dimensionality problem. In addition, the performance of the C-D control strategy faces great challenges of environmental uncertainty of user demand and electricity price. This paper proposes an EV cluster scheduling strategy considering real-time electricity prices based on deep reinforcement learning. Firstly, we establish a distributed real-time optimal scheduling structure according to the real-time price signals of distribution system operators (DSO). Furthermore, to alleviate the curse of dimensionality, we propose a C-D model of a single EV according to the C-D characteristics of EV, and we establish the C-D control model of EVs as a Markov decision process (MDP). Finally, to adapt to the uncertainty of the learning environment, we propose a model-based deep reinforcement learning to optimize the C-D behavior of EVs. After day-ahead training and parameter saving of the proposed model, the C-D scheduling strategy is generated for the real-time system state at each moment of the day. The simulation results of the C-D scheduling strategy for cost-oriented EV charging show that the proposed scheduling strategy effectively reduces the user charging cost by 133.7 dollars and the load peak-valley difference, stabilizes the load fluctuation, and achieves the win-win situation between the power grid and EV users. (C) 2022 The Authors. Published by Elsevier Ltd.