The time that a vehicle merges in a lane reduction can significantly affect passengers’ safety, comfort, and energy consumption, which can, in turn, affect the global adoption of autonomous electric vehicles. In this regard, this paper analyzes how connected and automated vehicles should cooperatively drive to reduce energy consumption and improve traffic flow. Specifically, a model-free deep reinforcement learning approach is used to find the optimal driving behavior in the scenario in which two platoons are merging into one. Several metrics are analyzed, including the time of the merge, energy consumption, and jerk, etc. Numerical simulation results show that the proposed framework can reduce the energy consumed by up to 76.7%, and the average jerk can be decreased by up to 50%, all by only changing the cooperative merge behavior. The present findings are essential since reducing the jerk can decrease the longitudinal acceleration oscillations, enhance comfort and drivability, and improve the general acceptance of autonomous vehicle platooning as a new technology.

This paper presents a practical experiment for estimating the state-of-charge (SOC) of individual cells in a series-connected heterogeneous lithium-ion battery pack, where only the terminal voltage of the battery pack is measured. To deal with real-time computation constraints, the dense extended Kalman filter (DEKF) algorithm has been proposed in the literature, which has a significantly lower computational complexity compared to the regular extended Kalman filter for this specific estimation problem. This work supplements the existing work by conducting a real-world experiment to validate the performance of the DEKF. Specifically, experiments involving a battery pack of three cells connected in series were conducted, where the battery pack was discharged under a constant current load. A genetic algorithm was applied to identify missing model parameters, as well as tuning the DEKF for optimal convergence and accurate SOC estimation. Our experimental results confirm that the proposed DEKF accurately estimates the SOC of each cell regardless of the hardware limitations and uncertainty, making it suitable for low-cost, real-time battery management systems. In particular, the SOC estimation error can be kept well under 1% even if the initial estimate is far from the true SOC.

This article presents a novel graph neural network (GNN)-based surrogate model (GNN-SM) for predicting state evolution in reconfigurable battery packs. By leveraging graph-based representations of battery cell interconnections, the proposed approach addresses the unique challenge of estimating the imbalance in state-of-charge (SOC) and temperature of cells of a battery pack in dynamic battery configurations. Unlike conventional methods that focus on instantaneous state estimation, our GNN-SM predicts future SOC and temperature distributions by considering both current system state and switch configuration. The model architecture combines graph attention networks with pooling operations to effectively capture cell-to-cell interactions and battery pack-level dynamics. Numerical results under constant current and constant power discharge conditions demonstrate that GNN-SMsignificantly outperforms baseline feedforward neural network (FNN) and FNN-attention models, achieving up to 73.7% reduction in root-mean-square error for temperature imbalance prediction and 46% reduction for SOC imbalance prediction. Furthermore, the model provides a 1629-fold speedup over high fidelity physics-based simulator while maintaining mean absolute percentage errors below 2% for temperature and 8% for SOC predictions. The scalability of GNN-SM is further validated on a 100-cell reconfigurable battery pack, where the proposed approach achieves high accuracy despite being trained on an extremely small fraction of all possible configurations. Finally, robustness analysis under sensor noise conditions demonstrates that GNN-SM maintains reliable predictions even under high noise levels.

This paper proposes a reinforcement learning (RL)-based control framework for real-time state-of-charge (SOC) balancing in reconfigurable lithium-ion battery packs. Traditional battery management strategies, such as passive and active balancing, suffer from energy inefficiency and limited adaptability. Reconfigurable battery systems offer enhanced flexibility through dynamic switchable topologies. However, to fully exploit the potential of the battery pack, an intelligent control policy to determine the optimal battery configuration in real-time is needed, which is a difficult task due to the large optimization space. To address this challenge, the reconfiguration problem is formulated as a Markov Decision Process, and a policy optimization framework is proposed to guide switch decisions based on observed cell states, such as the SOC of the individual cells in addition to the current pack configuration. The proposed RL agent operates on discrete action space and is guided by a multi-objective reward function tailored for battery management application. To avoid the high cost associated with physical battery training, we construct a linear surrogate model for efficient simulation-based learning. The proposed RL controller is evaluated against rule-based and greedy-based baselines, achieving superior SOC balancing performance with reduced switching frequency and improved runtime. In addition, comparison to true optimal obtained by dynamic programming demonstrates that the proposed RL controller achieves competitive results with significantly lower computational cost.