Corona discharge is a common type of partial discharge (PD) in high-voltage electrical equipment, and its prolonged presence can lead to equipment failure. The PD process generates rich ultrasonic signals, which are often detected using acoustic sensors. Currently, there is a lack of mechanistic analysis regarding the coupled electrical-thermal-acoustic processes during the PD. Therefore, we first establish a coupled electro-thermal-acoustic multi-physics simulation model, which reveals how space charge disturbances initiate acoustic waves and govern their propagation. Simulations show that the acoustic energy concentrates primarily within the ultrasonic band of 20 kHz to 100 kHz. Based on the primary frequency band distribution of the acoustic signal, we design an enhanced broadband acoustic focusing structure that integrates an acoustic Fresnel lens with a focusing cavity. A hybrid optimization algorithm combining Efficient Global Optimization (EGO) for global search and Nelder-Mead (NM) for local refinement is employed to optimize the parameters of the enhanced structure. Finally, the acoustic focusing performance of the optimization results is verified experimentally, achieving an average sound pressure amplification of 138.2% and a peak amplification of 272.8%. Additionally, Practical needle-tip discharge experiments verify that the introduction of the enhanced structure provide a relative enhancement in captured acoustic energy of approximately 48.99%. This study provides both theoretical underpinning and engineering application support for the acoustic detection of PD.

To investigate the impact of long-term voltage stress on the surface charge accumulation and discharge characteristics of insulating materials, Epoxy/Al2O3 composites were subjected to DC and AC voltage application tests for durations of seven and 30days. By integrating surface charge measurements, isothermal surface potential decay (ISPD) analysis, and scanning electron microscopy (SEM), the regulatory mechanisms of surface charge accumulation and trap distribution parameters on discharge behavior were analyzed. The results indicate that under long-term DC voltage, continuous unipolar charge injection induces microdefects such as grooves and pores on the material surface, leading to a significant increase in deep trap density. This enhancement in surface charge binding capability results in an increase in charge density with prolonged voltage application, which subsequently intensifies electric field distortion and partial discharge (PD) activity. On the contrary, the surface charge density under AC voltage is generally lower and exhibits an increasing trend from the high-voltage electrode toward the grounded electrode. Notably, the sample stressed for sevendays exhibited the lowest surface charge density yet the most intense PD activity. This phenomenon is attributed to an increase in shallow trap density, which enhances surface conductivity and carrier mobility. Conversely, after 30days of stress, the formation of deep traps restores the charge binding capability. Furthermore, the significant accumulation of negative charges near the grounded electrode under AC conditions is identified as a critical factor potentially triggering surface discharge.

With the increasing deployment of smart grids, non-intrusive load monitoring (NILM) has emerged as a crucial technology for estimating power consumption of individual devices using only aggregated measurements from the main power meter. Although federated learning (FL) offers a privacy-preserving solution for NILM by enabling collaborative model training across distributed smart meters, current FL-based NILM methods face two critical challenges: vulnerability to Byzantine attacks and inefficient local power disaggregation. This paper introduces the federated watermark closed-loop learning (FWM-CL) method to address these issues. Using inherent model redundancy, FWM-CL employs stable parameters as watermarks to detect and counter Byzantine attacks, significantly improving the robustness of the system. Furthermore, it integrates a feedback mechanism to optimise local power disaggregation network, achieving high accuracy with reduced computational complexity. Extensive evaluations on the UK-DALE and REDD datasets demonstrate that FWM-CL outperforms state-of-the-art methods in both Byzantine attack resistance and power disaggregation performance.

Objective Transformers are critical assets in power transmission and distribution networks, ensuring reliable electricity delivery and overall system stability. As transformers age, their failure probability increases, leading to higher maintenance costs and outage risks. Effective maintenance planning is therefore essential for sustaining reliability, extending service life, and mitigating failures. However, limited maintenance resources make the efficient scheduling of transformer inspections and maintenance a major challenge in asset management. Traditional reliability-centered maintenance approaches, which rely on historical data and risk matrices, focus on system-level reliability while often overlooking the individual operational characteristics of transformers. Moreover, most existing strategies optimize a single objective, without achieving a systematic balance between maintenance cost and failure risk. Methods To address these issues, this study proposes a cluster-based maintenance scheduling framework that explicitly considers asset heterogeneity and optimizes the trade-off between economic efficiency and reliability. The methodology integrates three components. First, a modified transformer failure rate model is developed by incorporating health index-based adjustments into a Weibull distribution, enabling individualized reliability assessments. The health index, derived from condition-monitoring data such as dissolved gas analysis indicators, provides a normalized, comprehensive condition score for each transformer. Second, the adjusted failure probabilities support asset-specific risk evaluation, allowing prioritized maintenance within each equipment cluster. The core decision variables define maintenance schedulesu2014specifying when each asset is taken offline and servicedu2014while adhering to operational feasibility and utility constraints, including failure rate thresholds, health index limits, allowable maintenance windows, and resource restrictions. Third, a dual-objective optimization model, formulated as a mixed-integer linear programming problem, determines the optimal timing and sequencing of maintenance tasks. Adjustable weight parameters enable flexible trade-offs between minimizing maintenance cost and reducing failure risk. Results The proposed approach was validated through this real-world case study, where simulation results showed a 12.6% reduction in total maintenance costs and an 8.2% decrease in average equipment failure risk compared with conventional methods. In addition, to analyze the impact of the maintenance coefficient on the optimization results, simulations were conducted using different coefficient values within a reasonable range while keeping other parameters constant. The results showed that larger maintenance coefficients led to poorer post-maintenance recovery, accelerated degradation, and an increase in average failure rate. Consequently, more maintenance actions were required to sustain system reliability, resulting in higher total costs. Moreover, by adjusting the reliability weight in the objective function while keeping other parameters unchanged, this study found that higher reliability weights corresponded to lower failure rates. When the reliability weight was set to 0.5, the model achieved the optimal balance between failure risk and maintenance cost, whereas overly low weights tended to maintain only the minimum acceptable maintenance intensity. Conclusions This study presents a comprehensive, data-driven maintenance strategy that integrates Weibull-based degradation modeling, health-index-adjusted failure prediction, and optimization-based scheduling. The flexibility of the maintenance scheme is also influenced by the scale of substation assets. When the maintenance coefficient changes, the optimization strategy may remain unchanged for smaller substations due to limited equipment quantities. In addition, the reliability weight can partially affect the optimized maintenance schedule, and tuning this parameter within a reasonable range allows utilities to obtain cost-minimized solutions while maintaining the desired reliability level. The proposed framework effectively balances reliability and economic efficiency, visualizes system-wide failure trends, and supports informed decision-making for substation asset management.

Gas-insulated switchgear (GIS) plays a critical role in high-voltage power systems, where structural defects, such as bolt loosening or base dislocation, can lead to catastrophic failures if not detected early. Traditional deep learning methods, including convolutional neural networks (CNNs), have achieved good performance in fault classification by extracting features from vibration spectrograms or time-frequency representations. However, these approaches lack physical interpretability, making it difficult to trace the causes of faults or generalize across different structural configurations. To address this limitation, a novel fault diagnosis framework that integrates physics-informed neural networks (PINNs) with an attention-enhanced shallow neural network (AESNN) is proposed, which first performs parameter inversion via PINNs to extract physically meaningful quantities. These quantities comprise equivalent damping, stiffness, strain energy, and the stress concentration factor, and are derived from limited vibration measurements. These parameters are then used as features to classify both fault types and severity levels in GIS systems. The results demonstrate that the proposed framework achieves superior accuracy when benchmarked against a portfolio of state-of-the-art deep learning paradigms. Furthermore, the physics-informed approach provides exceptional fault classification performance across diverse operational and structural conditions, along with a high degree of physical interpretability lacking in conventional black-box models. Furthermore, ablation experiments validate the importance of multiparameter coupling and the attention mechanism in improving classification performance. This study introduces a physically grounded alternative to black-box diagnosis models and offers a scalable methodology for intelligent diagnostics in complex electromechanical systems.

Moisture at the interface between the cross-linked polyethylene (XLPE) insulation and the silicone rubber (SiR) in the intermediate joint is one of the main causes of surface flashover in cable intermediate joints. However, research on the incipient fault caused by moisture at the intermediate joint interface is currently limited. This paper conducts incipient fault experiments on 10 kV cable intermediate joints exposed to moisture in a waterlogged environment. Based on power frequency voltage and current detection, a high-frequency current (HFCT) detection method is introduced to extract characteristic information, such as pulse amplitude and discharge frequency, at different stages of incipient fault development. The results show that the development process of the incipient fault at cable intermediate joints can be divided into three stages: initial, middle, and late stages. In the initial stage, the fault current amplitude is quite small, approximately 10–20 mA, and the high-frequency pulses are sparse and low in amplitude. In the middle stage, the fault current amplitude increases to 300 mA-850 mA, and the high-frequency pulse amplitude and frequency increase significantly. In the late stage, the fault current, high-frequency pulse amplitude, and discharge frequency all increase slightly compared to the middle stage. This research finding can provide a reference for early warning of distribution cable line faults.

The accurate measurement of internal mechanical parameters in gas-insulated switchgear (GIS) is essential for vibration modeling and transparent condition monitoring. Equivalent damping and stiffness are fundamental measurands that govern attenuation and resonance; however, in practical operation, they are difficult to obtain through direct experimental measurement, which poses a major challenge for online monitoring. This article proposes a physics-informed inverse measurement framework that estimates these parameters from sparse surface vibration signals. A high-frequency multiscale neural network is constructed to mitigate spectral bias and capture both low- and high-frequency dynamics, while geometry-aware modeling based on signed distance functions (SDFs) enforces boundary conformity in complex GIS structures. The training objective combines PDE-consistent residuals with boundary/initial constraints, ensuring the physical consistency without reliance on dense instrumentation. Validation on a GIS vibration platform with a 12-point accelerometer array and multiple excitation layouts shows a mean relative error of 3.97%, outperforming a baseline PINN (5.96%) and an SIREN (6.59%), and maintaining accuracy under 5% Gaussian noise (5.64%). The results indicate that the method operates as a virtual sensor for inverse measurement under sparse sensing, providing a traceable and robust metrological pathway for high-frequency vibration-based monitoring of high-voltage equipment.

Detection of SF6 Decomposition Components in Gas-Insulated Switchgear (GIS) is of significant importance for fault diagnosis. While infrared spectroscopy is a commonly used gas detection method with good stability and fastresponse, it still faces challenges such as unstable light sources, low resolution, and insufficient interaction between the gas and the optical path. This study employs an optical frequency comb (OFC) light source centered near 1550 nm. Leveraging its uniform frequency spacing and high-resolution characteristics, the source effectively enhances the detection _ accuracy of gas absorption spectra. An enhanced long-path gas absorption cell was designed to achieve fifty round trip reflections of the light beam within the chamber, resulting in an effective optical path length of 50 m, which was further validated by Monte Carlo ray-tracing simul ns. This design significantly improv he gas-lig tti n effect and consequently increases the gas absorption rate. By preparing CO and H2S gas mixtures at various concentrations, absorption spectrum measurements w d for bothssdifferent c els.The co sorbance

With the rising demand for Electric Vehicles (EVs), a coordinated Demand Response (DR) enabled framework is required to maintain grid stability and optimize energy consumption. Optimizing the charging demand at the Charging Station (CS) by managing resources is key to effective energy management. In this paper, we propose the Federated Learning Constrained Deep Reinforcement Learning (FLCDRL) framework, where each station acts as a local agent to optimize the schedule by following a local safety layer and periodically updates the central aggregator for the updated global schedule. For pricing, dynamic, market-driven coordination between agents is developed for energy bids and flexibility responses from/to the station and the Distributed System Operator (DSO) levels. Furthermore, the feeder voltages are managed by the safety layer in the Reinforcement Learning (RL) model, enabling the local agent to learn autonomously and contribute to global consensus through a voltage-balancing policy. Simulation results demonstrate that the federated TD3 agent effectively shifts EV demands to off-peak hours, minimizing operational costs while maintaining grid security. The proposed privacy-preserving framework validates multi-station EV charging scheduling, improving State of Charge (SOC) satisfaction by approximately 8.3% compared to other methods while preserving voltage limits and ensuring fair provisioning across different scenarios. It exhibits robustness and fair allocation by utilising unmet, Time Of Use (TOU), and federated model comparisons, ensuring compliance with SOC demand by EVs. The future framework focuses on uncertainty-aware coordination with Renewable Energy Sources (RES) and dynamic pricing mechanisms for resilient operation.

The increasing heterogeneity and uncertainty of power transformer degradation mechanisms in large-scale power grids have exposed the limitations of traditional maintenance strategies, which typically rely on predefined degradation models or centralized coordination assumptions. To address these challenges, this paper proposes an adaptive transformer maintenance decision-making framework based on a degradation–maintenance interacting Markov game and independent proximal policy optimization (IPPO) solution algorithm. The proposed framework explicitly captures how degradation and maintenance actions jointly shape asset health by modeling them as two interacting agents under decentralized decision-making framework, enabling adaptive maintenance decisions under uncertain degradation dynamics. An IPPO-based multi-agent reinforcement learning algorithm is further employed, allowing each agent to learn adaptive action policies independently according to its own objective while interacting through the shared environment. Case studies conducted on large-scale real-world transformer datasets demonstrate that the proposed method improves average episodic reward by over 80% compared with existing learning algorithms, while achieving a 14.9% reduction in total maintenance cost and a 22.2% improvement in overall cost-effectiveness relative to traditional RCM maintenance strategies.

Partial discharge (PD) inside gas-insulated switchgear (GIS) poses significant threats to the stable operation of electrical equipment, thereby rendering sensitive and stable PD detection critically important. Recently, the optical PD detection methods have garnered increasing attention due to their immunity to electromagnetic and acoustic interference. However, these methods still suffer from limitations, such as low capture efficiency for PD optical signal within GIS and inadequate detection sensitivity. Therefore, this article introduces a novel, high-sensitivity optical PD sensor based on a lightguide rod (LGR), whose sidewall is uniformly coated with CdSe/ZnS core-shell quantum dot (QD) material. Comprehensive demonstrations from theoretical analysis, optical simulation, and performance verification affirm that sidewall QD coating substantially improves the LGR's ability to capture PD optical signals, thereby improving PD detection sensitivity. Finally, PD experiments are conducted on an experimental platform built based on a small GIS tank. The QD-coated LGR showed higher response intensity and lower PD inception voltage (PDIV) in detecting four typical PD defects than traditional uncoated LGRs, achieving an average response gain of 37.4%. Experimental results indicate that the QD-coated LGR has considerable potential for practical engineering applications.

High-voltage direct current (HVdc) transmission technology has become a focal point in power system research. However, its development is constrained by the issue of surface charge accumulation, which leads to a decline in flashover performance. To investigate this issue, an experimental platform for surface discharge was established in this study. The surface charge density and flashover voltage of samples under different voltage durations were systematically tested. Through the integration of numerical simulation models and microscopic surface morphology analysis using scanning electron microscopy (SEM), this article systematically investigates the underlying mechanisms by which surface charge affects the surface flashover performance under long-term voltage application. The results show that surface charge accumulation, on the one hand, induces intensified electric field distortion, thereby reducing flashover voltage. On the other hand, under prolonged voltage application, plasma-induced surface modifications impede charge carrier migration, consequently enhancing flashover performance. For samples subjected to a voltage duration of seven days, the effects of the two factors are comparable. However, as the voltage duration increases to 30 days, the influence of the former becomes significantly dominant, leading to a marked decline in flashover performance.

Defects and hidden dangers of power equipment can be detected in advance by detecting the partial discharge (PD), and equipment failures and unexpected power outages could be avoided. Therefore, accurate orientation for PD is very important. However, time-delay-based PD electromagnetic signal orientation methods mainly rely on high-frequency sampling systems and high-precision synchronization performance, resulting in high hardware costs and relatively heavy devices. The existing methods based on the amplitude of ultrahigh-frequency (UHF) signal could only be used for 2-D orientation, which is hard to meet the application needs in actual 3-D field situations. Therefore, this article proposed an RSSI-based 3-D PD orientation method for electromagnetic signals, combining the visual image to generate a WYSIWYG detection of PD. First, based on the directional characteristics of the UHF sensor, a regular dodecahedron array structure was designed by simplifying a complex spherical structure. Then, a 3-D orientation model based on the amplitude of UHF signals was established, which included four variables: elevation angle, azimuth angle, maximum amplitude, and attenuation rate. Three-dimensional orientation results were obtained by solving the equation set. The cluster analysis was next employed for multiple 3-D orientation results to obtain the final accurate result. Finally, the accuracy of the proposed method was verified with experimental tests, and then, combined with the video, a PD imaging system was mounted on the UAV for a field test to verify the effectiveness of the actual defect.

Moisture and contamination ingress lead to a prominent failure issue of the epoxy-casting current transformer (CT)-air-insulating barrier (IB) system in 40.5 kV indoor switchgears. To address this, mechanisms of insulation failure from microscale material aging to macroscale system breakdown were investigated. Epoxy resin (EP) and sheet moulding compound (SMC) samples of CTs and IBs were prepared, and tests of surface resistivity, dielectric properties, contamination level, microscopic morphology, and molecular structure were conducted. For both CTs and IBs, the results indicate that the surface resistivity decreased by over 99.9 %, dielectric properties significantly deteriorated, and microstructural defects were observed. Raman spectra revealed breakages of C-H, C-O and C=C, ester bond hydrolysis, and existence of amorphous carbon. Primary aging mechanisms included structural degradation induced by synergistic erosion of moisture and contamination, organic material carbonization caused by discharge, and accelerated insulation failure driven by aging byproducts. Additionally, impacts of increased surface conductivity and barrier displacement on electric field strength and current density were investigated through finite element simulation. The results show that the formation of weakly insulating layers on material surfaces due to aging was essential for partial discharge initiation and elevated current density. The most probable interphase flashover path was identified as initiating from one busbar, propagating along the CT surface, then through air to the IB surface, continuing along the IB surface, then through air to another CT surface, and finally reaching another busbar, which aligned well with practical experience. The risk of interphase flashover in the CT-air-IB system increased with the increment of barrier displacement and was high when the conductivity of weakly insulating layers ranged from 0.001 to 0.01 S/m.

Transformer voiceprint recognition, which involves the analysis of transformer sounds to detect potential faults, has become a research hot spot in recent years. However, sound signals are susceptible to noise interference, and obtaining enough labeled transformer sound samples is challenging, thereby severely impacting the recognition accuracy. To this end, a deep learning algorithm for power transformer voiceprint recognition in strong-noise and small-sample scenarios is offered in this article. First, based on the frequency distribution characteristics of transformer body sounds, an improved spectrum method is proposed, selectively extracting transformer body frequency while suppressing environmental noise frequency. Second, ResVGG transformer voiceprint classification network is presented, which decouples training and inference to simultaneously enhance training accuracy and inference speed. Furthermore, a denoising network is introduced as a precursor to the classification network to further eliminate noise from the spectrum. Finally, self-supervised contrastive learning is employed, utilizing a large number of unlabeled sound samples for pretraining to enhance small-sample generalization and noise robustness. Experimental results demonstrate that the proposed method achieves a 10.6% improvement in accuracy compared to the baseline method, while also exhibiting satisfactory performance in inference time and memory usage.

The types of partial discharge (PD) in gas-insulated switchgear (GIS) are closely associated with the severity of insulation faults. Thus, making precise identification of PD patterns is crucial for assessing the operational status of GIS. However, there are two challenges in the existing research. First, single-modal PD exhibits limitations in detection capabilities, leading to difficulties in obtaining high-quality samples under specific conditions. Second, the correlations among samples, which reflect the shared features among samples of a specific PD type, are commonly neglected. To address the aforementioned two challenges, this article proposes a sparse cross-modal enhancement-driven multi-modal graph fusion network (SCE-MGFN). It can leverage auxiliary modalities to enhance the graph features of the dominant modality. The SCE module utilizes a sparse representation, thereby reducing computational complexity and pattern confusion risk during the multi-modal enhancement process. Additionally, a novel graph sample construction approach is proposed to extract the hidden spatial features of original phase-resolved PD (PRPD) patterns. The algorithm is validated on the full-scale GIS equipment. The model achieves the recognition accuracy of 97.5% on dataset D1 and 97.93% on dataset D2. The multimodal fusion method provides better theoretical foundation for future research.

Accurate simulation of vibration signals is essential for fault detection in power equipment such as gas-insulated switchgear (GIS) and transformers. The Finite Element Method (FEM) is commonly employed for high-fidelity simulations, but its precision heavily depends on exact physical parameters such as the damping coefficient. These parameters are difficult to measure directly, and existing methods based on empirical values are both time-consuming and often inaccurate. This paper proposes a method using Physics-Informed Neural Networks (PINNs) combined with experimental data to accurately invert damping parameters in vibration systems, exemplified by GIS. PINNs integrate physical laws into neural networks, improving accuracy, robustness, and generalization. By combining experimental data with PINNs, high precision and interpretability of key physical parameters in FEM simulations are achieved. The results show that without noise, the waveform similarity between the FEM and the experimental results is high, with an amplitude similarity coefficient of 0.869 and a normalized cross-correlation of 0.926. At the 1% noise level, the inversion error of the damping parameter is only 3%, and the method shows good noise resistance up to the 5% noise level. This approach improves simulation reliability and provides a new path to enhance the transparency and diagnostic capabilities of power equipment.

Non-Intrusive Load Monitoring (NILM) estimates load-specific power by disaggregating household-level power data, enabling smart grids to provide more accurate power estimations and thus prevent energy waste and casualties. Some existing NILM methods employ federated learning (FL) with generative models to estimate load power; however, their accuracy often suffers within an FL architecture. This is because the generators tend to learn the most common load patterns while neglecting the less frequent ones. To address this, we propose an FL architecture with a Wasserstein generative adversarial network (FL-WGAN) to enhance accuracy. In our method, each client trains its own generative neural network to estimate load power, while a discriminator network evaluates these estimates. Each client employs a Wasserstein distance-based guidance mechanism to ensure the generative model learns the full distribution of all states rather than being confined to a subset. Additionally, an attention mechanism is integrated into the generative model to further improve its representational capability. We evaluate FL-WGAN using the UA-DALE and REDD datasets, and the results demonstrate that our method outperforms existing methods.

High-voltage bushings are critical components in power transmission systems to maintain grid reliability. However, their complex electro-thermal-mechanical coupling and sealed structures pose significant challenges for internal state sensing and fault diagnosis. Traditional methods such as finite-element method (FEM) and partial discharge (PD) analysis often struggle to reconstruct internal stress and temperature fields due to sparse measurements, limited sensor accessibility, and the inherent ill-posedness of inverse problems. To address these issues, we propose a novel measurement method based on the physics-informed and experiment-guided multifield reconstruction network (PEG-MFRN), validated on a 110-kV high-voltage bushing. This framework integrates external discrete sensor data with governing physical equations to achieve high-fidelity reconstruction of internal stress and temperature fields, thereby enabling enhanced internal state transparency in the sense of reconstructing internal physical conditions from sparse external observations. Validation results demonstrate that the PEG-MFRN outperforms conventional FEM simulations, reducing maximum stress and temperature errors by 69.7% and 59.4%, respectively, and mean errors by 70.9% and 68.0%. The framework also exhibits robust performance under 5% Gaussian noise, with stress reconstruction found to be more sensitive to noise than temperature. Additionally, in coupled noise scenarios, the PEG-MFRN yields lower overall errors compared to single-noise cases, benefiting from its joint optimization and adaptive residual compensation mechanisms. These results demonstrate the potential of the PEG-MFRN as a noninvasive and physics-informed measurement framework and provide a reliable basis for transparent condition monitoring in power equipment.