Fault tracing of devices in transformer substations has attracted significant attention. Existing methods primarily rely on manually defined logic or features. And other statistical approaches lack consideration of the connection topology between devices. To address these limitations, this paper designs heterogeneous graphs for fault events, namely event graphs. It integrates the connection topology of devices with alarm signal information. The alarm signal information is transmitted by device monitors during fault events. In addition, an automatic graph classification framework is established to do the fault tracing, using the constructed event graphs. Within this framework, we propose a Hetero-Pooling technique to exploit node label information inherent in heterogeneous graphs. Furthermore, we incorporate the Graph-of-Graphs technique to alleviate the class imbalance issue prevalent in fault tracing data. Experimental results demonstrate that the proposed framework obtains good performance on real-world transformer substation dataset. It also performed well on some commonly used heterogeneous graph classification datasets from other fields.

The increasing integration of renewable energy and Electric Vehicles (EVs) introduces significant uncertainties into electricity markets, necessitating advanced participation strategies to cope with the risks arising from fluctuations caused by various factors. Existing studies have focused on optimization problems at hourly or daily timescales, often adopting bidding strategies designed with a single decision variable. This paper proposes a novel collaborative optimization framework for Virtual Power Plants (VPPs) equipped with Energy Storage System (ESS), which integrates risk-aware learning into the joint optimization of market bidding decisions and ESS dispatch strategies. Specifically, three key innovations are introduced. First, a three-segment quantity-price bidding strategy is developed to enable VPPs to submit differentiated bids based on real-time market prices and predefined psychological price ranges. Second, a risk-constrained Markov Decision Process (MDP) model is formulated, incorporating novel risk preference indicators and deviation penalty constraints to balance profit maximization with uncertainty mitigation. Third, a Worst-Case Risk considering Proximal Policy Optimization (WCRP) method is proposed, which employs a clipped operation to approximate the upper bound of Conditional Value-at-Risk (CVaR) within the objective function, ensuring policy updates explicitly account for extreme risks. Extensive experiments are conducted using historical data from the PJM U.S. electricity market covering 2024 Q2 and Q3. Our method reduces CVaR by approximate to 40% relative to a no-indicator ablation and increases mean profit by 3.0%; Also, our method attains the highest profit with the lowest CVaR and penalty, improving profit by 8.8-10.4% and cutting CVaR by 69.3-72.5% relative to benchmark methods.

This paper focuses on fault tracing and tracing rule mining of transformer substation devices. Current industrial methods use manually defined features or tracing rules. In addition, they need interpretability and consideration of the connection topology of devices. As an attempt to respond to these needs, this paper proposes a framework to model the substation device monitoring data into graphs, using them to trace faults and mine fault tracing rules. Specifically, this paper proposes a modeling method to construct event graphs based on the connectivity topological structure of devices and alarm signals. Then it addresses fault tracing as a graph classification task operated on the designed event graphs. To do this, this paper proposes a Weighted Generalized Weisfeiler-Lehman graph kernel (W-GWL), integrating a variant of the Weisfeiler-Lehman graph kernel with fault tracing domain expert knowledge. Meanwhile, this paper proposes an association analysis-based rule mining method to extract fault tracing rules, which is operated on Weisfeiler-Lehman subtrees. The mined rules can be easily interpreted by domain experts, which enhances the interpretability of the framework. Experiments were conducted on real-world transformer substation data, and the proposed W-GWL method performed well on precision, recall, and f1-score. In addition, the mined fault tracing rules match domain experts' logic. (c) 2025 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.

With integration of an energy storage system (ESS), an energy storage charging station serves as pivotal intermediaries between the smart grid and electric vehicles (EVs). This station utilizes the ESS to enhance grid stability and facilitate energy management. Participation in electricity market transactions offers revenue opportunities for charging stations, but it also introduces operational challenges, due to fluctuating electricity market prices and diverse energy demands and supplies. In this paper, we study the operation strategy optimization problem for the charging station, addressing economic and service challenges influenced by market volatility and energy diversity. The optimization objective considers not only maximizing economic benefits from the electricity market and EV services but also minimizing penalties associated with EV service quality. We propose a model that accounts for the dynamics of the electricity market, uncertainties from EV demands, and disturbances from green power generation, optimizing the power scheduling of the ESS and multiple charging piles (CPs) to determine transaction power in the market. The cooperative scheduling strategies for the ESS and CPs are learned using the proposed heterogeneous Multi-agent Deep Deterministic Policy Gradient method. This approach features distributed agents learning to determine decision variables for both the ESS and CPs, while a joint critic network assesses the station's overall objectives to guide their cooperative learning. The proposed method was tested against three state-of-the-art benchmark methods, which showed our method achieves better results.

Accurate prediction of lithium-ion battery cycle life is crucial for battery management systems and safety applications. This study presents a novel tri-branch fusion network (TBFN) for early-stage battery life prediction using the first 100 charge-discharge cycles. The methodology employs a comprehensive multi-dimensional feature extraction framework and three specialized branches to process different feature types: a Temporal Convolutional Network for sequential features, a Transformer for global dependencies, and a Multi-Layer Perceptron for statistical features. An adaptive gating fusion mechanism integrates the branch outputs to generate final predictions. Experimental validation on the MIT battery dataset demonstrates superior performance with 9.5% MAPE and 113.4 cycles RMSE, showing significant improvement over the original MIT study. The TBFN model exhibits robust prediction capability across diverse battery degradation patterns.

The power generation and loads of photovoltaic (PV) systems are inherently intermittent, periodic, and stochastic, resulting in nonstationary dynamic fluctuations in line currents. These dynamics have overlapping spectral features with dc series arcs, which increases the likelihood of false alarms and reduces the accuracy of arc detection. To address this problem, a dc series arc diagnosis method based on asymmetrical Gini impurity is proposed to reduce the false alarm rate (FAR) while ensuring accuracy. First, arc fault current data under various operating conditions are collected through experiments. To mitigate the impact of periodic fluctuations and enhance the detection of transient current changes, a multistep differential processing method is employed. Second, the time- and frequency-domain features of arc faults are extracted from the differential signal to construct a high-dimensional feature vector. To reduce redundant features, the clustering algorithm is applied to identify and group features demonstrating analogous classification performance, and the asymmetrical Gini impurity (AGI) is introduced to analyze feature significance. Finally, a random forest (RF) model based on AGI is utilized for dc series arc diagnosis. The experimental results demonstrate that the proposed method not only achieves high diagnosis accuracy and effectively reduces the FAR, but also completes fault detection within 60 ms.

Photovoltaic (PV) power generation exhibits significant stochasticity and volatility due to meteorological influences, challenging grid stability and energy management. This paper proposes a Spectral-Laplacian Enhanced Inverted Transformer (SLEIT) for accurate PV power forecasting. The architecture incorporates three key innovations: (1) a dual-domain fusion stack that adaptively integrates time-frequency features through parallel Fourier and temporal processing paths; (2) an inverted embedding mechanism that redirects attention to variable dimensions, enhancing multivariate dependency modeling; (3) a Laplace stack that approximates differential properties through cascaded nonlinear transformations, improving sensitivity to transient fluctuations. Experiments on six years of PV station data across multiple prediction horizons demonstrate that SLEIT achieves 3.63%-15.84% reduction in MSE and 0.43%-14.13% reduction in MAE compared to state-of-the-art baselines, with superior performance during extreme weather-induced power variations.

With the increasing demand for energy storage charging stations, many energy storage systems utilize lithium batteries as the major carriers. However, due to frequent charging and discharging at high power levels, the cycle life of lithium batteries is greatly reduced, which increases the energy storage costs. Given the longevity of supercapacitors, a supercapacitor-lithium hybrid energy storage system has been developed to effectively extend the lifespan of lithium batteries and reduce both investment and operational costs of energy storage charging stations. Based on the dual-stage active topology, a hybrid energy storage system combining supercapacitor-lithium is proposed. Under mild load conditions, two supercapacitor modules are alternatively charged by the lithium battery. Then, the supercapacitor modules are discharges when high power demands are encountered. Accordingly, based on working conditions of the charging pile, a multi-stage strategy, integrating state-of-power estimation and programming, is proposed to optimize the power distribution, smooth the power fluctuation of the lithium battery, and protect the lithium battery. The simulation results show that compared with the lithium batteries only energy storage and the traditional full active topology energy storage, the dual-stage active topology energy storage significantly improves the cycle life of lithium batteries.

Accurate estimation of the state of charge (SOC) is vital for the safe operation of battery systems. Traditional SOC estimation methods encounter challenges with LiFePO4 batteries due to their flat open circuit voltage characteristic in the middle range of SOC. Therefore., this paper proposes a closed-loop SOC estimation algorithm which improves the self-attention mechanism to build a neural network for preliminary SOC estimation and uses a particle filter to adaptively fuse the outputs of the neural network and the ampere-hour counting method for the final estimation results. The algorithm also provides a 95% confidence interval for the estimate., facilitating uncertainty assessment. Experimental results show that the proposed algorithm exhibits good performance for different temperatures and different cells within a battery pack., with a mean absolute error of less than 0.9% across the entire SOC range. Two extreme test cases demonstrate the algorithm's strong robustness to unknown initial SOC., sensor anomalies., and noise.

Nonuniform thermal behavior in lithium-ion battery packs can accelerate aging, leading to inconsistent cell performance. If not adequately monitored and managed, this heating can give rise to unwanted side reactions, fires, and explosions, underscoring the criticality of temperature field reconstruction. In recent years, data-driven methods have gained popularity for addressing the temperature field reconstruction problem. However, many existing data-driven approaches require retraining when system parameters change, such as the initial temperature distribution or working conditions. This article presents a deep transfer operator learning method named physics-informed adversarial networks. The model architecture incorporates transformer blocks to capture comprehensive time and space features. Additionally, to enhance interpretability and generalization, the model introduces two effective mechanisms: 1) the integration of thermal partial differential equations to ensure compliance with physical laws; and 2) the application of domain adversarial mechanism in transfer learning to extract domain-invariant feature representations. These mechanisms enable the model to effectively reconstruct the temperature field, even in unencountered scenarios during training. The proposed method is validated under real-world energy storage working conditions, demonstrating superior performance compared to state-of-the-art deep learning methods. Notably, the approach exhibits excellent performance even when confronted with the limited availability of training data.

The grid load forecasting plays a crucial role in the steady operation management of smart grid system. However., the drastic and irregular fluctuations of grid load make the accurate forecasting difficult to achieve, and most of the existing models are unable to learn the complex data patterns of load data. In this work, we propose a novel framework named Neural Frequency Interpolation Model (NFIM) to tackle these challenges. NFIM is composed of 3 stacks, with each stack using a pooling layer and a multilayer perceptron (MLP) to extract feature representations from the input sequence, which are then decoded by different basis decoder modules. Such a structure helps the model learn the deep patterns of grid load data, such as pattern in frequency domain and multi-scale resolution, thereby achieving high-performance forecasting. Extensive experimental results on real-world dataset verifies that our model outperforms state-of-art deep learning models on grid load forecasting tasks.

In this work, the second-order consensus in discrete-time multi-agent system (MAS) with reference states in which only a part of agents have access to is researched. According to the proposed network topology, it is unnecessary to utilize a spanning tree in the directed graph. First, three different protocols are adopted to demonstrate three typical cases on reference states. Then, necessary sufficient conditions to achieve the required consensus with reference states are derived in MASs. At last, several examples are calculated to verify the proposed theory.

Due to the wide zero-voltage-switching range and low power losses, triple-phase-shift (TPS) modulation is commonly utilized in dual-active-bridge (DAB) converters. However, it is difficult to model it and design its controller for the reasons of model uncertainties and nonlinearity. In this paper, a deep reinforcement learning (DRL)-supervised proportional–integral (PI) control algorithm is proposed. The PI controller is used as a base controller to stabilize the output voltage of the DAB converter. In order to improve the control accuracy and the dynamic performance, the PI parameters are tuned by DRL. Besides, all operation modes of the TPS are learned during the training process. Thus, the operation mode with maximum power efficiency can be selected under a wide operation range. The simulation comparison results demonstrate the efficacy and superiorities of the proposed method.

To reduce the peak power caused by fast charging of numerous electric vehicles, and to decrease the cost of fast charging stations, a hybrid energy storage system composed of super capacitors and lithium batteries, corresponding to high power density devices and high energy density devices, respectively, is developed to improve the economic benefit of charging stations and alleviate the load fluctuation by properly allocating the capacity of energy storage. A mixed integer nonlinear model is built to evaluate the optimal configuration of the hybrid energy storage system, by minimizing the total cost of the fast charging station and establishing the real-time power balance, along with the working state of the energy storage system as constraints. A two-stage cooperative optimization algorithm is proposed to solve this model. The simulation results demonstrate that the peak power can be significantly reduced, and the total cost of the fast charging station can be effectively reduced, which validates the feasibility of our proposed capacity configuration model and algorithm.

Heat generation significantly influences the performance of lithium-ion batteries and also hinders the application of them. Precise prediction of battery temperature can offer feedbacks to monitor system so as to enable safe and efficient operation of batteries. However, battery temperature prediction remains extremely challenging due to the increase of irreversible heat caused by aging across the life cycle. To tackle this problem, we propose a novel framework named battery informed neural network (BINN). In this paper, we incorporate battery physical models into long short-term memory (LSTM)-based networks that is trainable in an end -to-end manner for battery temperature prediction. Multi-head attention mechanism is introduced to attend to information from longer time series. Physical parameters of the battery electrical model, heat generation model, and thermal model are automatically learnt during training. The irreversible heat changes brought on by aging is considered and represented by physical parameters. Temperature prediction for the full life cycle of lithium-ion batteries using BINN is tested under different working conditions. It is shown that BINN is interpretable and has better generalization and transferability than traditional learning-based methods.

随着人们对生活品质要求的提高,混纺粗毛纱作为高端毛纺织品在市场上的需求量逐年攀升.然而,混纺粗毛纱的生产工艺和机械性能影响仍然是制约其产业发展的关键因素之一.本研究通过对混纺粘胶、涤纶、锦纶和羊毛的粗毛纱制作过程中加捻程度的调整,探究了加捻对纱线断裂强度和断裂伸长率的影响.结果表明,在一定的加捻程度范围内,纱线的断裂强度与断裂伸长均随着加捻程度的增加而呈现先上升后下降的趋势.加捻过程中,粗毛纱中的纤维同时收到拉伸和摩擦作用,适当的加捻能提升粗毛纱的机械强度,同时受混纺纤维的种类与性能影响显著.本研究结果可为混纺粗毛纱生产工艺的优化提供理论依据和实践指导.

The assessment of battery health has long been a major concern in lithium-ion battery applications. Effective and efficient approaches for the estimation of state-of-health (SOH) are crucial to LiFePO 4 batteries with poor consistency. In this paper, we propose an estimation method based on Gaussian process regression (GPR) algorithm using partial charging curve. The curve is within a voltage range of only 60 mV, which is determined by a pre-analysis of voltage cumulative distribution. Considering the grade from grey relation analysis and the complexity to obtain, three interpretable health indicators are extracted from the charging voltage sequence as inputs to the model with a combined kernel function. The model achieves a root mean square error of 0.88% and outperforms several benchmark models on the tested public dataset. Furthermore, experimental results using data from the established platform reveal the high accuracy and adaptability of the proposed method.

This paper deals with the sliding mode stabilization for chaotic systems. In the system under consideration, the nonlinear function is one-sided Lipschitz with quadratic inner-boundedness. Specifically, a non-fragile sliding mode surface is constructed, and the sufficient condition for the convergence is derived. Then, a new feedback law is proposed to enable the state trajectories of the closed-loop system to reach the sliding mode surface in finite time. Finally, an example in the background of the unified chaos system is simulated to show the validation of the designed controller.

A fabric quality detection method based on the YOLO architecture combined with the attention mechanism is proposed. A computer vision aided data acquisition system collects images for defect detection. The goal is to create clean samples for on-linear detection. The proposed method outperforms traditional flaw detection methods such as statistical analysis or heuristic algorithms in terms of accuracy and detection speed. A convolution neural network model is used to create the deep learning model, which detects defect features using cascade convolution layers and cross stage partial network structures as the backbone network. The leaky rectified linear unit and residual units are used to effectively determine and extract features. block attention module is used to improve detection accuracy even further. Fabric flaw quality detection can thus be kept effective and precise. For performance validation, an open source practically captured data set of fabric with and without flaws is used.

为实现数控加工中直线转角处的连续平滑进给,提出了一类直线转角的非对称平滑加工规划方法.针对转角初末速度、加速度大小不同的情况,考虑直线转角处的运动学约束和几何误差约束,并将高次转角模型求解化简为线性规划问题,可直接解算转角参数,实现进给速率平滑与刀具轨迹平滑的同步规划,提高规划效率与加工效率.与对称转角平滑的仿真对比实验结果表明,所提方法在保证加工过程平稳的情况下效率与精度有所提高,运动平台样机实验验证了该方法的可行性.