The digital twin (DT) is a promising technology that will provide a cyber-physical platform to enable rapid product development boosted by artificial intelligence and big data. This paper focuses on DT-based optimization of the energy management strategy (EMS) for a plug-in hybrid vehicle (PHEV). A dedicated adaptive particle swarm optimization (DAPSO) algorithm is developed to enhance the optimality and trustworthiness of the DT-based EMS optimization in a variety of driving conditions. By using the chassis dynamometer data, the DAPSO is developed by incorporating the widely-used PSO algorithm with an adaptive swarm control strategy that coordinates the exploration and exploitation of the individual swarming particles following the global information extracted from the swarm. With the cross-validation testing under three worldwide driving conditions, this study determines the unified settings of the DAPSO algorithm for the PHEV application. Experimental evaluations are conducted by monitoring the PHEV’s performance using different EMSs with control thresholds optimized by the DAPSO and conventional PSO algorithms (baseline). The results show that by introducing DAPSO for DT-based EMS optimization, up to 24% improvement in cost function value (weighted sum of the fuel consumption and SoC sustaining error) and more than 49.1% reduction in computing time can be achieved compared to the baseline methods.

The fuel cell electric vehicle (FCEV) is a promising pathway for transport decarbonization. Managing the FCEV’s onboard energy is a challenging task that requires advanced control of multiple power sources. The equivalent consumption minimization strategy (ECMS) has shown effectiveness in the energy management of hybrid electric vehicles (HEVs), and its performance depends on the settings of the equivalent factor (EF). This article proposed a meta-heuristic fuzzy inference ECMS method (MA-ECMS) for FCEV to improve efficiency and robustness. The fuzzy inference system (FIS) of the MA-ECMS is optimized with a meta-heuristic algorithm, chaotic-enhanced particle swarm optimization (CAPSO), which optimizes the EF settings adaptively based on battery state-of-charge (SoC) and vehicle power demand. A PI-based SoC penalty regulator is added to the MA-ECMS to improve the SoC tracking capability. The optimality and robustness are evaluated for different driving cycles, and initial battery SoC states using the ECMS and A-ECMS as baselines. The real-time performance of MA-ECMS is verified on a processor-in-the-loop (PiL) testing platform. The results suggest that the MA-ECMS can achieve lower hydrogen consumption, better robustness, and fewer SoC tracking errors. Compared with the baselines, up to 13.13% hydrogen can be saved.

The energy management strategy (EMS) of the fuel cell electric vehicle (FCEV) is a control process that crucially affects the FCEV’s economic performance and driving range. Conventionally, EMS optimization relies on standard driving cycles, such as the worldwide harmonized light vehicles test cycle (WLTC). However, these often lack robustness due to their limited representation of real-world scenarios. This study proposes an adversary evolutionary learning (AEL) scheme that generates challenging driving cases during the optimization process to enhance robustness. This study demonstrated the effectiveness of AEL based on an FCEV controlled by the equivalent consumption minimization strategy (ECMS), where the equivalent factor (EF) settings must be fine-tuned. The AEL scheme involves interactive attack and defense rounds. In the attack round, a genetic algorithm (GA)-based cycle generator creates challenging driving cycles to stress-test the EF settings. In the defense round, particle swarm optimization (PSO) tunes the EF set by a fuzzy inference system (FIS) for energy efficiency and stability in the state of charge (SoC) of the battery. After 30 rounds, AEL identifies 30 harsh driving cycles and 30 optimized ECMS. Crossover testing is then performed to select the most robust ECMS (R-ECMS) set. Processor-in-the-loop (PiL) experiments on standard real-world cycles demonstrate that AEL effectively identifies the R-ECMS, potentially saving up to 1.37% of hydrogen and reducing SoC fluctuations by up to 47.43%.

To guarantee the safety of electric vehicles (EVs), it is crucial to gain a comprehensive understanding of internal short circuits (ISCs) and to develop effective early detection methods within battery management systems (BMSs). This study presents a detailed analysis of ISC detection in cylindrical lithium-ion battery (LIB) cells, utilizing a mechanistic coupled electrochemical–thermal model to replicate various ISC scenarios. The proposed approach integrates an extended Kalman filter (EKF) with a coupled lumped-parameter electrothermal model, enabling real-time ISC detection leveraging both electrical and thermal measurements. Extensive evaluations demonstrate that the coupled model significantly improves detection accuracy and timeliness, reducing detection times by 36% and errors by 44% on average compared to uncoupled models in different ISC scenarios. These findings are particularly significant for preventing with thermal runaway (TR), especially in the presence of severe ISC. Robustness analysis further highlights the impact of ISC location on detection accuracy. The study concludes that the proposed ISC detection framework offers an efficient, accurate, and reliable solution to detect ISC faults in cylindrical battery cells, contributing to safer operation. By mitigating TR risks and extending battery lifespan, this work supports the development of sustainable energy systems and improves the reliability of energy storage technologies.

This article develops a robust energy management method for plug-in hybrid electric vehicles (PHEVs) by proposing an adversary swarm learning (ASL) system. The proposed ASL system includes an attacking swarm and a defending swarm working adversarial on an iteration base to obtain the most robust equivalent consumption minimization strategy (ECMS) for PHEV energy management. In the attacking rounds, the defender fixes the ECMS settings while the attacker trains a model to generate the worst case driving profiles that aim to obtain the maximum equivalent fuel and electricity consumption. In the defending rounds, the defender optimizes the setting of ECMS in the driving cycle generated by the attacker. By introducing two new concepts, attacking rate and defending rate, the robust settings for ECMS were obtained using the ASL systems based on the minimum attacking rate (MAR) and the maximum defending rate (MDR), respectively. Using the conventional ECMS optimized under four standard driving cycles as the baseline, this article demonstrated the advantage of the proposed methods on both the software-in-the-loop and hardware-in-the-loop platforms. The study suggests that the robust ECMS obtained by the proposed ASL systems outperforms the baseline in all selected driving cycles, and the ASL using MAR strategy is shown more effective by achieving an average cost-reducing rate of 10.58% compared to the baseline.

There is a trend to implement end-to-end learning-based control for electric vehicles; however, how to ensure the optimality and safety of the learning-based control methods is a great challenge. This article proposes a novel ensemble Q approach incorporating in-target minimization and update-to-data (UTD) mechanisms to control a plug-in hybrid vehicle that has an end-to-end control architecture. By employing these mechanisms, the prevalent issue of Q-function overestimation bias in Q-learning can be addressed with a reduction from 1.18 to 5.2 $\times $ $10^{-4}$ . Meanwhile, the proposed method can also improve sample efficiency, outperforming conventional Q-learning methods in terms of learning stability and performance. The advantages of the proposed method in terms of energy efficiency and learning stability were validated through experiments on software-in-the-loop (SiL) and hardware-in-the-loop (HiL) platforms. The results show that up to 10.09% and 15.59% energy savings can be achieved through the comparison with conventional Q-learning and double Q approaches.