Mirror-symmetrical Dijkstra’s algorithm-based deep reinforcement learning for dynamic wireless charging navigation of electric vehicles

Abstract

The dynamic wireless charging (DWC) system based on wireless charging lanes (WCLs) is an important component of smart cities, allowing electric vehicles (EVs) to charge while moving. It is necessary to establish a user-oriented real-time DWC navigation system to achieve the joint optimization of EV routing and charging. However, the modeling characteristics of DWC and the risk preferences of EV owners towards congested WCLs are completely different from those in traditional wired charging. Furthermore, optimal EV charging navigation is always challenging without prior knowledge of uncertainty in electricity prices and traffic conditions. This paper first proposes a novel dynamic charging routing model for individual EVs to minimize travel and charging costs, and reformulates it as a twostep optimization problem to facilitate feature extraction. Then, mirror-symmetrical Dijkstras algorithm (MSDA) is proposed to solve the reformulated model in linear time and extract advanced features from the stochastic information. By feeding the system state containing extracted features into the deep Q network (DQN) in an event-triggered manner, the near-optimal charging navigation strategy is finally obtained. The proposed MSDADQN approach not only efficiently extracts low-dimensional interpretable input features, but also adaptively learns the unknown dynamics of system uncertainty. Numerical results based on simulated and real-world data validate the proposed approach.