Joint optimization of charging station selection and dynamic charging price: A predictive deep reinforcement learning approach

The rapid adoption of Electric Vehicles (EVs) presents critical operational challenges, including extended waiting times at charging stations (CSs), congestion-induced inefficiencies, and uneven power demand across the distribution grid. Existing methods often fall short in addressing these complexities holistically, particularly where user autonomy and network-level constraints intersect. To bridge this gap, a novel bilevel decision-making framework is proposed, combining real-time station recommendation and adaptive pricing. The Lower-Level (LL) formulation incorporates user-centric factors such as State-of-Charge (SoC), travel distance, dynamic pricing, and queue length, while the Upper-Level (UL) accounts for temporal grid characteristics including weekday/weekend classification, office-hour sensitivity, and peak-load mitigation. The methodology integrates a Multi Deep Regression (MDR) module for arrival rate forecasting with a hybrid Deep Reinforcement Learning (DRL)-based control layer, where a Soft Actor-Critic (SAC) network optimizes charging prices and a Dueling Deep Q-Network (DDQN) directs users to optimal CS locations. The interaction between these components creates a feedback loop that dynamically aligns user behavior with system objectives. Simulation results, validated using real-world datasets from California, demonstrate a 25.5 (%) reduction in congestion, 2.8 (%) decrease in average waiting time, and a 26.8 (%) shift in peak load demand to off-peak periods. The smart agent's reward trajectory outperforms human-based strategies by 18.8 (%), achieving learning accuracy ( R 2 ) near 99 (%). These outcomes confirm the model’s effectiveness in balancing grid efficiency with user satisfaction, without compromising data privacy or behavioral realism.