System configuration represents the core decision stage in system design and the physical basis for dynamic power distribution. This study develops a bi-level optimization framework to address multi-timescale coupling in fuel cell hybrid power system configuration for unmanned aerial vehicles. Under multiple performance constraints, the upper-level system configuration layer employs the self-organizing surrogate-assisted non-dominated sorting differential evolution algorithm to balance hydrogen utilization efficiency, lightweight design, and endurance requirements while ensuring operational safety and fault tolerance. The lower-level energy management layer applies dynamic programming to evaluate upper-level solutions. Configuration results for the DJI T10 prototype indicate a fuel cell rated power of 2.5 kW, battery energy of 395.16 Wh, and hydrogen tank volume of 12 L. Comparative experiments demonstrate the framework’s effectiveness in reducing computational burden while maintaining applicability to various multirotor unmanned aerial vehicle configurations and mission requirements, exhibiting strong generalization capability. Applicability boundary analysis confirms that fuel cell hybrid power systems are well-suited for heavy-load, long-endurance missions under both current and future technology scenarios.

With the rapid development of the new energy vehicle industry, the consistency of power battery packs has gradually become a key factor affecting system safety and lifespan. Addressing inconsistencies caused by variations in voltage, capacity, and internal resistance during long-term operation, this paper proposes a lightweight diagnostic method integrating attention mechanisms with a multi-layer perceptron (SE-Transformer-MLP). Based on real-vehicle operational data, this method extracts four-dimensional features—peak positions and peak values of primary and secondary peaks—through preprocessing of charging segments and construction of Incremental Capacity (IC) curves. It incorporates State of Health (SOH) as a fifth-dimensional feature input. By integrating electrochemical mechanisms and expert knowledge, inconsistencies are categorized into three root causes, establishing a supervised learning labeling system. The model utilizes the SE module to highlight key features, employs the Transformer encoder module to capture global dependencies, and achieves multi-class classification through the MLP module. Experimental results demonstrate that the proposed SE-Transformer-MLP model achieves high classification accuracy on the test set, exhibiting superior feature discrimination capability and robustness compared to traditional shallow models. This provides an effective approach for diagnosing root causes of battery inconsistencies. The research findings hold significant reference value for health assessment and root cause diagnosis of inconsistencies in battery management systems.

The design of charging strategies for lithium-ion batteries in electric vehicles typically demands extensive experimentation and large datasets to model the battery behavior. This paper presents a hybrid charging optimization framework that combines data-driven learning with physics-based modeling. The degradation mechanism of lithium-ion batteries is analyzed and incorporated as physics-informed features to guide the charging strategy. A multi-layer perceptron is employed to predict capacity loss using minimal historical data, providing high adaptability to nonlinear battery behaviors. The charging control involved multiple objectives, including battery aging, temperature raise and time constraint. The constrained optimization by linear approximations optimization algorithm is utilized to determine optimal charging currents under both safe and fast charging scenarios. Its capability to handle non-linear constraints and black-box objective functions ensures robust multiple objectives optimization across varying conditions. Experimental results demonstrate that the proposed method effectively reduces charging time and extends battery life, while requiring minimal training data. Through experiments and simulations, the proposed charging control method achieved a 20.7% and 88.2% improvement in reducing SOH degradation during charging as compared with traditional CCCV methods.

Abstract External short circuits (ESCs) in lithium-ion batteries can cause rapid temperature rise and safety issues. This study proposes an online EIS-based method for diagnosing and assessing the severity of ESC faults. By integrating real-time electrochemical impedance spectroscopy (EIS) measurements with characteristics of external short circuits, the approach enables dynamic fault evaluation. Prior to a fault, EIS data are used to fit a fractional-order model and extract parameters. During an ESC event, voltage–current data are fitted via nonlinear least-squares to a double-exponential model to update key parameters, while time-domain features including peak temperature and voltage drop are simultaneously extracted. These parameters and features are fed into a regression model that predicts fault duration as a severity indicator. An ESC evaluation index is then derived and classified into three severity levels: mild, moderate, and severe. Experimental results demonstrate that the method achieves high accuracy and clear severity grading, proving suitable for real-time ESC assessment in battery management systems.

Prognostic and health management (PHM) can well compensate for the defects of proton exchange membrane fuel cell (PEMFC) stacks in terms of safety and reliability. However, related techniques are limited by the pre-diagnosis accuracy of multiple fault modes. To address the above challenges, this paper proposes an integrated framework that synergizes dynamic state prediction with fault classification to pre-diagnose balance of plant (BOP) faults. This method realizes the association rule mining between different sensor features, and uses them as physical constraints to construct the dynamic behavior error (DBE) loss function. This mechanism enhances the model’s ability to capture the long-term dynamic evolution of faults. Finally, the real-time similarity between feature behaviors and rules is used to assist fault pre-diagnosis. The results show that compared to the latest methods, the proposed framework improves the accuracy of pre-diagnosing 9 BOP faults by approximately 10%, while reducing single inference time by at least 43.3%. The proposed framework provides a novel solution for predictive diagnosis of hydrogen-recirculation stack faults under anode purging disturbances, and its universality is verified on different models.

With advancing transportation electrification, electric vehicles (EV) have become key parts of urban power consumption. Their charging demand shows spatiotemporal uncertainty, which brings severe challenges to power grid stability. This perspective critically reviews studies on EV-VPP integration. It focuses on three core areas: VPP dispatch, charging demand forecasting, and charging station recommendation. Current research faces three major limitations. They are fragmentation focused on individual component optimization, insufficient robustness under uncertainties, and inadequate analysis of real-world constraints. We propose the Multi-Scale & Multi-Agent Collaborative Robust Framework (MSMRF) as a viable solution for this research field. Centered on uncertainty, the framework establishes a closed-loop collaborative mechanism of "forecasting-dispatch-recommendation". It also integrates various practical constraints into robust algorithm design. This perspective deepens the understanding of EV-VPP integration. It notes that the key to large-scale integration lies in balancing multi-agent interests and improving overall system robustness. It should not limit optimization efforts to individual components. The critical analysis of existing technologies and the proposed innovative framework provide a new perspective. This perspective supports research on EV charging guidance under the VPP architecture.

Modern manufacturing is undergoing a transformation driven by the pursuit of efficiency, precision, and reliability in mechanical systems. As critical equipment, computer numerical control (CNC) lathes play an indispensable role, with their stability being paramount to preventing economic losses and production delays. This paper proposes a novel framework for CNC lathe fault diagnosis based on fine-tuned large language models (LLMs). The Hierarchical Supervised Fine-Tuning (HSFT) algorithm balances sequence generation, hierarchical classification, and Chain of Thought (CoT) consistency regularization losses by optimizing a composite loss function. This enhances model adaptability to precisely address domain-specific diagnostic challenges. The CoT inference module constructs a structured stepwise reasoning process, improving diagnostic accuracy and interpretability through multi-stage methods including fault classification, self-verification, and reasoning. Experimental results on Qwen and LLaMA series models demonstrate significant performance improvements over baseline models, validating the approach’s effectiveness. This framework provides a robust solution for practical CNC lathe fault diagnosis and delivers actionable diagnostic insights for maintenance engineers.

Accurately predicting the aging process of fuel cells and extending their lifespan through effective management strategies is crucial. This paper proposes a dual-channel prediction framework based on Adaptive Extended Kalman Filter (AEKF) and Random Forest (RF) models. The approach utilizes an AEKF algorithm, based on a semi-empirical aging model, to predict the irreversible aging trajectory of fuel cells under normal operating conditions. Simultaneously, a Random Forest model, optimized using the grey wolf optimizer, is employed to predict the reversible aging trajectory caused by improper operation. Specifically, the proposed method first applies the AEKF algorithm, derived from the semi-empirical aging model, to extract the baseline trend representing irreversible aging. Then, the dataset undergoes a detrending process to isolate the reversible aging component, which serves as the training datasets of RF model. This ensures that the RF model effectively captures the reversible aging characteristics of the fuel cell. The method is validated using experimental data from proton exchange membrane fuel cells under both constant and quasi-dynamic load conditions. Using 60% of the dataset under two different operating conditions, the proposed approach achieves a mean absolute percentage error of only 0.464% and 0.502%, respectively.

Safe and stable operation of distribution networks is crucial to power supply reliability. Automatic defect detection based on unmanned aerial vehicle (UAV) inspection images has become an important research direction for power operation and maintenance. However, aerial UAV images suffer from large-scale object variation, a high proportion of tiny defective objects, complex backgrounds, and low contrast, which impose higher requirements on detection accuracy and real-time performance. To address these challenges, this paper proposes an improved detection model based on BiFPN and BiFormer for RT-DETR. Taking RT-DETR-R18 as the baseline, the bidirectional feature pyramid network BiFPN is introduced to enhance multi-scale feature fusion capability. Meanwhile, the BiFormer bi-level routing attention mechanism is embedded to establish sparse global dependencies, which effectively improves feature discrimination and bounding box regression accuracy for low-contrast tiny defects. Experimental results demonstrate that compared with the original RT-DETR-R18 baseline, the proposed model achieves improvements of 3.56, 4.79, and 3.70 percentage points on mAP50, mAP75, and mAP50–95, respectively. The recall rate of hard-to-detect objects (Class B) increases by 3.18 percentage points, while maintaining a real-time detection speed of 70.53 FPS. The proposed method can well meet the practical application requirements of distribution network UAV inspection.

For heavy-duty commercial vehicles with multi-energy powertrains, the delayed response of key power sources makes predictive energy management difficult. Accurate power demand prediction is therefore essential for improving energy allocation and vehicle efficiency. Conventional methods often rely on vehicle speed forecasting, but multi-step speed prediction can introduce error propagation, which limits their effectiveness under highly dynamic and nonlinear heavy-duty driving conditions. To address this problem, this study develops a driving-cycle dataset for intelligent connected heavy-duty commercial vehicles using a SUMO-MATLAB co-simulation platform. Based on this dataset, a direct power prediction method is proposed using a cross-attention Long Short-Term Memory (LSTM) network. Historical power data are used as the main input channel to describe the vehicle's intrinsic power evolution, while intelligent transportation information, including Controller Area Network (CAN) data, Vehicle-to-Everything (V2X) data, perception data, and other variables related to the traffic environment and driving behavior, is introduced as an auxiliary channel. The two information streams are fused through cross-attention to improve prediction accuracy. The sparrow search algorithm is further used to optimize the network parameters. Simulation results show that the proposed method reduces the Mean Absolute Error (MAE) by 41.3%. Real-vehicle validation further confirms its effectiveness, with a 74.2% MAE reduction. When applied to the energy management of an ammonia-hydrogen hybrid heavy-duty vehicle, the proposed method reduces fuel consumption by 2.94%.

The variability of driving conditions and the complexity of traffic environment have brought great challenges to the energy management of fuel cell hybrid vehicles. Under the background of the rapid development of Internet of vehicles and intelligent transportation technology, a predictive hierarchical energy management framework for proton exchange membrane fuel cell/lithium-ion battery hybrid vehicles integrated with traffic perception speed prediction is proposed. Using traffic flow simulation and vehicle-to-everything (V2X) data flow, a short-term speed prediction model based on spatio-temporal graph neural network (STGNN) is established. By fusing historical vehicle data and real-time traffic information, the prediction accuracy is significantly improved. Based on the above speed prediction ability, a hierarchical model predictive control energy management strategy considering health is designed. The upper layer completes the planning of battery state of charge reference trajectory by using the long horizon vehicle speed preview information, and the lower layer realizes model predictive control, and carries out real-time multi-objective energy optimization for hydrogen consumption, component life and reference state of charge (SOC) tracking. The simulation results indicate that, in comparison to benchmark energy management strategies, the framework has significant advantages in extending the life of energy system and reducing operating costs.

Durability remains a critical barrier to the large-scale commercial adoption of proton exchange membrane fuel cells (PEMFCs). To address the limitations of pure data-driven methods in physical interpretability and model-driven approaches in prediction accuracy, this paper proposes a hybrid prediction framework integrating Crossformer with physics-informed learning for PEMFC degradation voltage forecasting. The framework employs convolutional layers and Crossformer's cross-dimensional attention mechanism to effectively extract local temporal features and model dependencies among multiple parameters, enabling accurate characterization of the dynamic voltage decay behavior. Moreover, by incorporating electrochemical mechanisms as physical constraints and adopting a three-stage progressive training strategy, it achieves synergistic optimization between data-driven learning and physical modeling. Experimental results on three aging datasets under steadystate, quasi-dynamic, and dynamic current conditions demonstrate that the proposed model achieves superior generalization performance across various operating scenarios. The model attains root mean square error (RMSE) values of 0.00191-0.00305 V and mean absolute percentage error (MAPE) values of 0.0339%-0.323%, outperforming Informer, convolutional neural network (CNN), and other comparative models. Ablation studies further confirm the individual contributions of the convolutional module and physical constraints to prediction accuracy, as well as the model's robustness under limited training data conditions.

Hybrid AC/DC microgrids with distributed energy storage (DS) improve power reliability in remote areas. Existing power management methods either focus on steady-state power sharing or transient inertia support, but rarely combine both. They also often ignore frequency and voltage deviations caused by droop control, which can harm sensitive loads. To overcome these issues, this paper proposes a full-time-scale (FTS) power management strategy that unifies transient inertia sharing and steady-state power allocation through a novel dynamic concatenator. It also introduces autonomous frequency/voltage restoration to eliminate steady-state deviations in each subgrid. Additionally, a global equivalent circuit model (GECM) is developed to simplify system analysis and design. Experiments confirm that the approach maintains nominal frequency and voltage in steady state while enabling seamless transition between transient inertia support and proportional power sharing across all time scales.

UAV-based inspection is the dominant approach for power grid defect detection, yet accurate identification of small-texture insulator discharge traces and structural line clamp insulation cover missing defects remains challenging. The state-of-the-art RT-DETR suffers from unsatisfactory fine-grained detection performance, as its AIFI encoder’s standard attention is insensitive to edge information and the feed-forward network lacks spatial inductive bias. To address this, we propose EA-AIFI, an edge-aware Transformer encoder module. It introduces a dual-scale differential edge-aware attention bias to focus on defect boundaries, and a spatial-aware ConvFFN to capture subtle texture and structural anomalies. Experiments on our private dataset show that EA-AIFI achieves 4.5% absolute mAP@0.5 improvement over baseline RT-DETR-R18 (0.609 vs. 0.564), with a remarkable 7.8% boost for insulator discharge traces, while maintaining real-time inference. With negligible overhead, it is highly suitable for edge deployment in industrial power inspection.

Accurate state-of-health estimation is critical for the safe and reliable operation of Lithium-ion batteries, but developing robust data-driven models is often hindered by data privacy concerns and the distributed nature of the data. Federated learning (FL) offers a privacy-preserving solution, yet its performance is severely challenged by the statistical heterogeneity and data imbalance inherent in real-world distributed battery data. Furthermore, the opaque nature of conventional FL models impedes the cross-institutional trust required for collaborative training. To address these limitations, this article proposes a novel multiphysics-informed adaptive personalized federated learning (MP-APFL) framework. At the client level, we first develop a multiphysics-informed, multistage aging partial differential equation (PDE) that combines analytical models for key degradation mechanisms, such as solid electrolyte interphase formation, thermal effects, and lithium plating-induced “knee points,” with a data-driven term to account for complex unmodeled dynamics. Subsequently, this governing PDE is embedded as a soft constraint within a multiphysics-informed neural network via an adaptively weighted composite loss function that synergistically enforces physical consistency, maintains parameter realism, and ensures trend plausibility by penalizing unphysical jumps while permitting valid recovery, guiding each local model to learn physically plausible and interpretable aging dynamics. At the system level, we design MP-APFL to employ a privacy-preserving hybrid architecture with a shared encoder and personalized multiphysics components. This is orchestrated by an adaptive strategy featuring Information value-driven client prioritization for efficient client selection and physics-guided and stability-aware encoder aggregation for robust model fusion. Comprehensive experiments on laboratory-generated and public datasets demonstrate that MP-APFL significantly outperforms prominent FL benchmarks, particularly in highly heterogeneous and imbalanced settings.

With the rise of electric vehicles (EVs), transportation electrification has become vital for sustainable mobility and low-carbon development. This study proposes a two-stage optimization framework for EV charging stations (CSs), combining user equilibrium (UE) modeling with multiagent reinforcement learning (MARL). A behavior-aware demand allocation model is first constructed based on real-world road networks and points of interest, capturing users' adaptive routing and charging responses under spatially heterogeneous pricing. Building upon this, a graph-enhanced MARL approach is employed to enable decentralized and competitive pricing decisions. By leveraging local graph structures, each CS agent perceives neighboring pricing and demand conditions to adapt strategies accordingly. Experimental results indicate that the proposed method increases revenue by 17.5% and effectively alleviates congestion and spatial imbalance, demonstrating its potential for coordinated optimization of transportation and power distribution systems.

This paper presents an integrated framework for energy-efficient trajectory planning and predictive energy management of quadrotor unmanned aerial vehicles (UAVs) powered by hydrogen fuel-cell–battery hybrid propulsion systems. The trajectory planning problem is formulated as a bi-objective optimisation that simultaneously minimises total propulsion energy and flight distance in cluttered three-dimensional environments. A preference-weighted NSGA-II algorithm is employed to generate a Pareto-optimal set of waypoint sequences, which are subsequently converted into smooth, dynamically feasible trajectories using arc-length-parameterised cubic splines that enforce speed, acceleration, and angular-rate constraints. [1]At the energy-management level, a Weighted Extended -Horizon Model Predictive Controller (WMPC) coordinates the power split between the fuel cell and battery by optimising a cost function that accounts for hydrogen consumption, fuel-cell degradation, and battery ageing. The controller combines an accurate short-horizon prediction with a computationally efficient extended look-ahead window using exponentially decaying weights. A physics-based battery equivalent-circuit model provides real-time SOC feedback.Simulation results demonstrate that the proposed co-design approach significantly reduces operational cost and component degradation compared with rule-based and standard MPC baselines, while satisfying all mission constraints.

Accurately predicting the capacity degradation trajectory of on-board lithium-ion batteries is crucial for alleviating range anxiety among electric vehicle users and ensuring the safe operation of vehicles. However, constrained by complex operating environments and stochastic driving behaviors, real-world vehicle data commonly face challenges such as high-intensity environmental noise interference, scarcity of full-lifecycle samples, and severe missing data, rendering existing data-driven methods difficult to apply directly. To address these issues, this paper proposes a battery capacity prediction method tailored for real-world vehicle data. First, to obtain reliable capacity labels from fragmented charging data, a capacity extraction method based on charging mode analysis and temperature correction is proposed. Furthermore, the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise algorithm combined with contribution analysis is utilized to effectively eliminate noise components and reconstruct the long-term degradation trend. Second, to mitigate the overfitting problem caused by the scarcity of real-world data, a Wasserstein Generative Adversarial Network with gradient penalty is introduced for data augmentation, generating high-fidelity virtual degradation trajectories to enrich the dataset. Finally, this paper innovatively leverages the continuous-time modeling characteristics of the Closed-form Continuous-time neural network to achieve precise prediction of capacity degradation trajectories under data missing conditions. Verification based on 29 months of real-world operation data from 20 commercial electric vehicles demonstrates that the proposed method achieves an RMSE of 0.6955 Ah, significantly reducing the prediction error by 17.7% compared to the optimal baseline algorithm. Furthermore, when facing severe data-missing conditions, the proposed method outperforms advanced time-series prediction algorithms by reducing the error by 36.8%, maintaining exceptional robustness and stability.

Low-temperature fast charging leads to lithium plating, posing severe threats to the safety and lifespan of lithium-ion batteries. Existing lithium plating detection methods often suffer from destructive testing requirements, high computational cost, or limited physical interpretability. This study proposes a physics-informed parameter identification approach for rapid and non-destructive lithium plating diagnosis. A Genetic Algorithm with Multi-Scale Dilated Convolution framework (GAMDIL) is developed, integrating genetic algorithms with multi-scale dilated convolutional neural networks to achieve efficient identification of key electrochemical parameters in the pseudo-two-dimensional model. Based on the identified parameters, a lithium plating diagnostic framework is constructed, which is oriented toward P2D parameter migration analysis and incorporates incremental capacity features. Experimental results demonstrate that the parameters identified by GAMDIL achieve a voltage error of approximately 15 mV under constant-current conditions and below 20 mV under dynamic operating conditions. Furthermore, the proposed lithium plating diagnosis framework correctly identified all tested cells, including independently prepared lithium-plating and non-plating validation samples, with results consistent with scanning electron microscopy observations, demonstrating its high diagnostic accuracy and robustness.