A modern battery management system (BMS) in electric vehicles plays a crucial role in enhancing battery pack safety, reliability, and performance, particularly in e-transportation applications. To achieve more accurate estimation methods, combining battery digital twinning with cloud computing for computational power and data storage capabilities proves beneficial. Over the last decade, various data-driven state-of-charge (SOC) estimation methods, such as machine learning and deep learning approaches, have been introduced to provide highly precise estimations. The widely used SOC estimation method in the industry is the extended Kalman filter (EKF). To explore and analyze the potential use of SOC estimation in a cloud platform, this paper develops and conducts a comparative analysis of four SOC estimation methods: EKF, Feedforward Neural Network (FNN), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM). These models are deployed in two cloud computing infrastructures, and their accuracy and computing time are thoroughly examined in this study. The study concludes that the EKF method is the fastest and most accurate among all considered methods. It boasts an average execution time of 54.8 ms and a mean absolute error of $\mathbf {2\times 10^{-4}}$ when measured over a physical distance of approximately 450 km via the mobile network Long-Term Evolution (LTE).

Health-conscious battery management systems (BMSs) that rely on surface temperature measurements are insufficient for managing automotive lithium-ion batteries (LIBs). Experimental studies have shown temperature differences of up to $10~^{\circ }$ C between surface and core of cylindrical LIBs. BMSs that consider only surface temperature overlook critical thermal information. The missing monitoring can delay detecting thermal events within the cell, accelerating battery degradation and increasing the risk of thermal runaway. This article introduces two deep learning algorithms to address this: Kolmogorov-Arnold network (KAN) and interconnected long short-term memory (LSTM) network. Both approaches estimate the core temperature of LIBs without requiring surface temperature feedback to the neural network. Experimental validation revealed a core temperature mean absolute error (MAE) of $0.55~^{\circ }$ C with a computational cost of 2.9–3.2 ms for KAN. The proposed interconnected LSTM reached a MAE of $0.80~^{\circ }$ C. The performance of the two core temperature estimation techniques was further evaluated under dynamic loading profile using urban dynamometer driving schedule (UDDS) drive cycle. The KAN method achieved a MAE of $0.325~^{\circ }$ C, demonstrating its adaptability to dynamic operating conditions. The two proposed methods, primarily KAN, are both adaptive and computationally efficient, making them suitable for integrating onboard BMS and cloud-enabled digital-twin-based BMS systems.

In the context of the electrification of the mobility sector, smart algorithms have to be developed to control battery packs. Smart and reconfigurable batteries are a promising alternative to conventional battery packs and offer new possibilities for operation and condition monitoring. This work proposes a reinforcement learning (RL) algorithm to balance the State of Charge (SoC) of reconfigurable batteries based on the topologies half-bridge and battery modular multilevel management (BM3). As an RL algorithm, Amortized Q-learning (AQL) is implemented, which enables the control of enormous numbers of possible configurations of the reconfigurable battery as well as the combination of classical controlling approaches and machine learning methods. This enhances the safety mechanisms during control. As a neural network of the AQL, a Feedforward Neuronal Network (FNN) is implemented consisting of three hidden layers. The experimental evaluation using a 12-cell hybrid cascaded multilevel converter illustrates the applicability of the method to balance the SoC and maintain the balanced state during discharge. The evaluation shows a 20.3% slower balancing process compared to a conventional approach. Nevertheless, AQL shows great potential for multiobjective optimizations and can be applied as an RL algorithm for control in power electronics.

Lithium-ion batteries are a key technology for the electrification of the transport sector and the corresponding move to renewable energy. It is vital to determine the condition of lithium-ion batteries at all times to optimize their operation. Because of the various loading conditions these batteries are subjected to and the complex structure of the electrochemical systems, it is not possible to directly measure their condition, including their state of charge. Instead, battery models are used to emulate their behavior. Data-driven models have become of increasing interest because they demonstrate high levels of accuracy with less development time; however, they are highly dependent on their database. To overcome this problem, in this paper, the use of a data augmentation method to improve the training of artificial neural networks is analyzed. A linear regression model, as well as a multilayer perceptron and a convolutional neural network, are trained with different amounts of artificial data to estimate the state of charge of a battery cell. All models are tested on real data to examine the applicability of the models in a real application. The lowest test error is obtained for the convolutional neural network, with a mean absolute error of 0.27%. The results highlight the potential of data-driven models and the potential to improve the training of these models using artificial data.

Towards smart batteries for electric vehicles (EVs) smart algorithms to control battery packs, mainly reconfigurable batteries, have to be developed. This work proposes a reinforcement learning (RL) algorithm to balance the State of Charge (SoC) of reconfigurable batteries based on the topologies half-bridge and battery modular multilevel management (BM3). As RL algorithm, Amortized Q-learning (AQL) is implemented, which allows enourmous numbers of possible configurations of the reconfigurable battery to be controlled, as well as the combination of classical controlling approach and machine learning methods. This enables safety mechanisms in control. As a neural network of the AQL a Feedforward Neuronal Network (FNN) is implemented consisting of three hidden layers. The experimental evaluation using a 12-cell hybrid cascaded multilevel converter illustrates the applicability of the method to balance the SoC and maintain the balanced state during discharge. The evaluation shows a 20.3% slower balancing compared to the classical. Nevertheless, AQL shows great potential to be applied for multiobjective optimizations as an applicable RL algorithm for control in power electronics.