A Deep Q-Network Framework for Joint Optimization of EV Charging Station Placement and Vehicle Routing

Abstract

The rapid adoption of electric vehicles (EVs) presents new challenges in designing efficient charging infrastructure and route planning mechanisms to ensure sustainable urban mobility. This paper introduces a novel two-stage optimization framework that leverages Deep Q-Networks (DQNs) for the intelligent placement of EV charging stations and real-time routing of vehicles. In the first stage, a DQN placement model learns to identify optimal charging station locations within a graph-represented road network, minimizing the expected energy consumption of future routing. In the second stage, a separate DQN routing model is trained to approximate cost-to-go functions and guide vehicles to the nearest stations efficiently. The models are evaluated against traditional methods, including Q-Learning, neural networks (NN), deep graph neural networks (DGNN), and Random Walk baselines. Simulation results across diverse network topologies demonstrate that our DQN approach consistently outperforms baselines in terms of average energy
consumption, travel time, and route success ratio. Specifically, the DQN placement strategy achieved the lowest average energy consumption of 1.2387 kWh and the most spatially equitable configuration (average distance: 6.34 km), while the DQN routing model recorded the best performance with 1.2587 kWh average energy, 1.98 hops, and a 96.67% success rate. These findings highlight the effectiveness of deep reinforcement learning in enhancing the scalability, reliability, and energy efficiency of EV infrastructure planning.