Chuanlin Zhang

With the increasing penetration of distributed photovoltaic (DPV) systems, active distribution networks (ADNs) are facing growing challenges such as frequent voltage violations and rising operational losses. The reuse of retired energy storage system (RESS) provides a cost-effective solution for voltage regulation. However, the constraints of state of charge (SOC) and state of health (SOH) impose stricter requirements on scheduling strategies. This paper proposes a multi-timescale voltage optimization method for ADNs considering RESS, where the operation time characteristics of multiple voltage regulation devices are incorporated to achieve hierarchical optimization of voltage deviations. Specifically, aiming to minimize voltage deviation and improve the coordination of device operations, a hybrid deep reinforcement learning (DRL) framework is developed. In detail, a Deep Q-Network (DQN) is employed for discrete device control based on value functions, and Deep Deterministic Policy Gradient (DDPG) is applied for continuous device action optimization via policy gradients, as well as Proximal Policy Optimization (PPO) with clipped policy updates is used to enhance RESS charging and discharging stability. Additionally, SOC depth and SOH degradation are explicitly incorporated into the reward function to limit battery aging while ensuring effective operation. Simulation results on the IEEE 33-bus test system demonstrate that the proposed approach achieves an average voltage deviation of 0.0164 p.u. and reduces daily network losses to 17.291 MW compared with benchmark methods, proving its effectiveness under scenarios with high PV output uncertainty. Over a 24-hour simulation, OLTC tap changes and SCB switching actions are both limited to 4 times. In addition, RESS power fluctuations remain within 500 kW, SOC is maintained between 0.2 and 0.8 respectively, and SOH degradation is reduced to 2% as relative slowdown value in specific high volatility testing scenarios, which effectively mitigating battery aging. These results verify the advancement and practical value of the proposed multi-timescale voltage optimization strategy.

Data-driven approaches have proven effective for lithium-ion battery state of health estimation. Nevertheless, their practical performance is frequently constrained by distribution shifts across different batteries and degradation phases, which undermine the model’s generalization capability under real world conditions. To address this issue, this paper proposes a Gaussian mixture weighted domain generalization network (GM-WDGN). Firstly, a Gaussian mixture model is employed to segment the source domain data into multiple subdomains with distinct temporal distributions, facilitating a more refined characterization of distribution heterogeneity. Subsequently, domain adversarial learning is introduced to achieve cross-domain distribution alignment, while a weighted smoothing label strategy is further incorporated to enhance training robustness. The resulting model can then be directly applied to estimation tasks on target domain samples. Experimental results indicate that the proposed approach improves estimation accuracy by 23.3% to 40.5% compared with conventional domain generalization methods. Furthermore, visualization results reveal a more consistent alignment of feature distributions across different domains.

Hand-eye calibration of line laser cameras (LLCs) is a prerequisite for accurate 3D measurement and reliable welding path execution in robotic welding systems. However, calibration methods based on planar or spherical targets often exhibit fitting ambiguity, sensitivity to illumination variations, and limited robustness under industrial operating conditions. To address these limitations, this paper presents a novel hand–eye calibration framework guided by the geometry of a triangular pyramid target for LLCs. By exploiting the surface normal discontinuities of the triangular pyramid, the proposed approach induces stable and unique laser extremum features during scanning, enabling consistent feature extraction and well-posed geometric calibration constraints. Under these constraints, the hand-eye transformation is formulated as a constrained optimization problem and solved using sequential least squares quadratic programming (SLSQP), which enforces rotational orthogonality and bounded translation. Experiments on a real robotic welding system show that the proposed method achieves a calibration error of 0.648 mm based on the bootstrap 95% confidence analysis, and maintains reliable performance across both comparative and field evaluations.

With the intensification of climate change and the accelerating shift toward transportation electrification, electric vehicles (EVs) play a critical role in enabling transportation decarbonization and sustainable mobility. As the EV population continues to grow, the management of user charging behavior has become increasingly important. In this context, high-quality charging station recommendations not only improve users’ charging experience but also help reduce unnecessary energy consumption, thereby supporting low-carbon mobility goals. However, traditional recommendation approaches, typically based on static optimization or conventional machine learning, struggle to capture the heterogeneity of user attributes and psychological preferences, limiting the effectiveness of personalized charging recommendations. To address this challenge, this study proposes EV-Rec-R1, a framework that integrates large language models (LLMs) with reinforcement learning (RL) through group relative policy optimization (GRPO). By incorporating user characteristics, psychological preferences, and charging station information, the framework constructs a behavior-aware charging station recommendation model. A composite reward mechanism guides the model toward generating more accurate and interpretable recommendations. Experiments conducted using real-world survey data and charging behavior records demonstrate that the proposed method significantly outperforms traditional machine learning models and unoptimized language models. The findings provide methodological insights into EV charging behavior, improve personalized charging services, and contribute to the development of intelligent and electrified transportation systems.

This paper presents a fuzzy sampled-data tracking control framework for a class of general nonlinear systems subject to unmatched disturbances, developed through a composite nonrecursive synthesis approach. First, focusing on unmatched disturbance rejection, an Euler discretization is applied to a continuous-time nonsmooth disturbance observer, resulting in a homogeneous sampled-data observer. Subsequently, by incorporating the estimated disturbances to modify the original system, a homogeneous nonrecursive sampled-data controller is directly derived via a simple coordinate transformation. In the semiglobal stability analysis, a relationship between the sampling period and the system bandwidth factor is established. Furthermore, to address the well-recognized influence of sampling period variations on system performance, a fuzzy logic criterion describing the performance-sampling period relationship is introduced. By this criterion, a fuzzy self-tuning mechanism is employed to more flexibly determine the most appropriate sampling period for the system under real-time operating conditions. Finally, the effectiveness of the proposed strategy is validated through numerical simulations and experimental studies on robot manipulators, demonstrating its practical application potential.

In numerous practical systems, the dynamics inherently involve multiple time scales. For example, in motors, power electronic, and other electrical circuits, fast electromagnetic transients coexist with slow mechanical motions. Motivated by this observation, this paper develops a two-timescale event-triggered control strategy tailored for such multi-timescale systems. To conserve communication resources, an efficient two-timescale dynamic event-triggered mechanism is designed, in which an auxiliary variable is introduced to dynamically adjust the triggering threshold, while a nonlinear function is introduced to further accelerate the convergence of the error system, balancing the system performance and bandwidth resource. On this basis, a two-timescale adaptive controller is proposed to adjust the network coupling strength, which depends on the slow-fast timescale parameters. Furthermore, by leveraging the characteristics of network topology, the consensus conditions are extended to a situation where computational complexity is independent of the number of network nodes, in which case the adaptive parameter is constant. Finally, to validate the effectiveness of the proposed method, two illustrative examples including an electric circuit network and five single-link manipulators are provided.