Schweitzer Engineering Laboratories

Voltage sags (dips) are transient reductions in RMS voltage $(\mathbf{1 0 - 9 0 \%}$ of nominal for $\mathbf{0. 5}$ cycles to $\mathbf{1}$ minute), represent one of the most critical power quality disturbances affecting industrial operations, causing equipment misoperation and significant economic losses. Accurate classification of voltage sags into the seven IEEE 1159 types (A-G) is essential for precise fault diagnosis, root cause analysis, and targeted mitigation strategies. However, traditional supervised learning approaches for voltage sag classification require extensive labeled datasets, which are expensive and time-consuming to obtain, limiting practical deployment in real world power systems. This paper introduces a semi-supervised learning framework combining FreeMatch with Vision Transformer architecture for label efficient voltage sag classification. By employing adaptive confidence thresholding and self adaptive fairness mechanisms, our approach effectively leverages unlabeled voltage sag waveform data while preventing class imbalance. Using synthetic three phase waveforms following IEEE 1159 and IEEE 1564 specifications, the method achieves competitive 84.14% accuracy with only 15% labeled data, approaching fully supervised performance while dramatically reducing annotation costs by 85%. This represents the first application of FreeMatch to voltage sag classification, specifically, demonstrating that advanced semi-supervised techniques enable practical deployment of fine-grained sag type identification in resource constrained industrial environments where expert labeling of power quality events remains a persistent bottleneck.

Successfully detecting high-impedance faults due to downed or broken power distribution conductors in a timely manner has been a big challenge for decades. When an energized broken conductor makes contact with the ground, it may result in a high-impedance fault that may be a challenge to detect using traditional protection methods. This paper provides a review of the existing solutions to detect downed conductors that have made contact with the ground. It is important to highlight that these solutions detect and isolate the affected circuit section only after the energized conductor has been on the ground for several seconds or minutes. This creates a critical “race-against-time” scenario, posing wildfire risks and public safety hazards. This paper dives deeper into an innovative method that was developed and successfully implemented on 12 kV distribution circuits to detect and isolate broken conductors while they are in the air and before they touch the ground. The IEC 61850 Generic Object-Oriented Substation Event and IEEE Std. C37.118 synchrophasor-based falling conductor protection solution is designed to detect and isolate broken conductors well within 500 ms of the break. This protection-speed solution is applicable to three-phase circuits along with two-phase and single-phase laterals that may be in high fire risk areas. This paper further explores the implementation of the falling conductor protection solution using Ethernet radios, and direct fiber, as well as private long-term evolution communication networks, which form the backbone of the falling conductor protection solution.

Modern Battery Management Systems (BMS) have a vital role to play in the accurate estimation of the State of Charge (SOC) in lithium-ion batteries, which has a direct impact on the safety, reliability, and performance. Although deep learning has proven to be promising, models trained using symmetric loss functions tend to have an overestimation bias. To measure and deal with this particular risk, this paper presents a specialized training approach for the LSTM model that was trained with five different loss functions and an evaluation framework that quantitatively measure the tendency of a model to make unsafe predictions. The experiment shows that there is a distinct trade-off between accuracy and safety. The Asymmetric Huber model is the most accurate, with the lowest Mean Absolute Error (MAE) of 0.0081. The soft safety margin model is the most optimal in terms of maximizing safety with a very low overestimation rate of 4.9 %. Finally, the Conservative Quantile model is selected as the best solution, which represents the best balance, having a good MAE of 0.0123 and the second-lowest rate of overestimation. These results demonstrate that risk-conscious, asymmetric loss functions engineering is a very efficient methodology towards creating safer and more credible SOC estimation models.