Assessing the aging of mineral insulating oil is essential for the dielectric reliability of oil-immersed power transformers. Terahertz time-domain spectroscopy (THz-TDS) non-destructively probes the dielectric changes caused by aging, but the measured spectra are also shifted by the initial moisture level of the oil, which limits the generalization of data-driven aging classifiers to an unseen moisture level. This paper proposes a physics-guided terahertz dielectric response learning (PG-TDRL) framework for assessing transformer oil aging across different initial moisture levels. The absorption coefficient and dielectric loss tangent are extracted from THz-TDS measurements as physically interpretable descriptors. A Fisher analysis of the training conditions alone identifies the terahertz bands that best separate the aging stages, and a physics prior layer softly emphasizes them. A spectral encoder and a prototype classifier are further regularized by a sample-level contrastive alignment loss that reduces the feature drift caused by the moisture condition. Under a leave-one-condition-out protocol on 336 THz-TDS measurements from 42 aged oil samples, PG-TDRL achieves 94.6% accuracy and 94.5% macro-F1, outperforming representative domain generalization methods. The results indicate that PG-TDRL makes terahertz aging assessment of transformer oil more robust to the initial moisture level.
This study investigates the cable length induced distortion in polarization and depolarization current (PDC) based insulation diagnosis and proposes a correction method to improve the reliability of characteristic parameter extraction for long cables. PDC measurements are first performed on cables with different lengths. The DC conductivity and tanδ are calculated and compared. The results show that tanδ is significantly increasing as the cable length increases, indicating that cable length introduces a non-negligible distortion into the extracted diagnostic parameters. To analyze this phenomenon, the polarization and depolarization processes of cables with different lengths are studied using an equivalent circuit model of PDC measurement. The corresponding diagnostic parameters are calculated and compared with the actual insulation response. It is found that when the cable length reaches to the point where the cable capacitive reactance is less than or equal to the the current-limiting resistor, a significant deviation appears in the tanδ derived from PDC measurement. This deviation is further interpreted using the Extended Debye Model, which clarifies the mechanism of cable-length-induced distortion in PDC characteristics. To eliminate this effect, a correction method based on singular value decomposition (SVD) is proposed. The proposed method is applied to long submarine cables, and the corrected results verify its validity and provide a reference for promoting the field application of the PDC measument.
Sudden flashovers of polluted and wetted line insulation lead to significant power outages. For this reason, monitoring the pollution degree of insulation is very important. This research develops a novel methodology for remote assessment of pollution degree on high-voltage external insulation using the analysis of surface partial discharge radiation. Experimental investigations establish that spectral properties of surface discharges exhibit systematic and distinct variation, correlated with pollution levels: from predominantly blue-violet emissions on clean surfaces toward yellow-orange tones on polluted insulators. A quantitative color ratio of surface discharge radiation demonstrates a significant linear correlation with directly measured surface conductivity and enables the non-invasive determination of insulation pollution level, timely washing and cleaning of insulators when necessary.
The valve-side main insulation of converter transformers is simultaneously subjected to power-frequency AC, DC, and harmonic electric fields. The electric field distortion caused by charge behavior under complex electric fields makes partial discharge (PD) more likely to occur. Based on the measured voltage waveform under the low-valve-hall Y/Δ connection of a converter transformer, the PD and charge characteristics of oil-pressboard insulation under needle-plate electrodes are experimentally investigated. This AC/DC superimposed harmonic voltage is hereafter referred to as the Y/Δ voltage. Combined with discharge simulation capable of mapping experimental charge data, the effect of voltage waveform steepness on PD is mainly studied from the perspective of charge. The results show that the density at the peak energy-level of the deep traps in the pressboard under the Y/Δ voltage is 19.32×109 m-3·eV-1. The maximum electric-field distortion rate reaches 33.6 %, which is 45.5 % higher than that under the AC/DC composite voltage with a ratio of effective values of 1:1. The electric-field distortion induced by space charge accumulation causes some PD pulses to appear on the falling edge of the voltage. Simulations show that high waveform steepness causes multiple stagnation events during the lateral evolution of the discharge streamer, thereby increasing the average discharge magnitude and producing a multi-peak discharge distribution under the Y/Δ voltage. The results provide theoretical support for condition assessment and modification control of converter transformer oil-pressboard insulation.
Transformer oil is a critical insulating and cooling medium in oil-immersed power equipment, and its degradation directly affects insulation reliability and service life. Breakdown voltage is a key indicator of transformer oil dielectric strength; however, conventional measurements are time-consuming and destructive. This study proposes an ultrasonic pulse-echo-based approach for estimating transformer oil breakdown voltage by integrating time–frequency acoustic features with a data-driven regression model. An ultrasonic detection platform was developed, and thermally aged 25# Karamay transformer oil samples were prepared under controlled conditions. A total of 162 time- and frequency-domain features were extracted from ultrasonic echo signals, among which 82 strongly correlated with breakdown voltage were selected using the maximal information coefficient. Based on these features, a support vector regression model optimized by the dung beetle optimization algorithm was established. Experimental results show that the proposed method achieves good estimation accuracy and repeatability against standardized breakdown voltage measurements. This study provides a non-destructive and efficient approach for transformer oil condition assessment, with potential applicability to different oil types through appropriate model training or recalibration using representative data from the corresponding oil. However, the physical mechanisms underlying the correspondence between ultrasonic acoustic features and breakdown voltage remain to be further clarified.
