DOI: 10.11648/j.epes.20261505.11 ISSN: 2326-9200

Digital-twin and Physics-informed Neural Ensemble for Locating Asymmetrical Short-circuit Faults on 10 kV Overhead Feeders

Ramziddin o‘g‘li, Nurmatov Yoqubboyevich
Show PDF Cite
This paper proposes a digital-twin-assisted, physics-informed (feature-enriched) ensemble for locating asymmetrical short-circuit faults on a 10 kV radial overhead feeder using single-ended current and voltage measurements. The digital twin randomizes feeder sequence impedances, source impedance, loading, fault location, fault resistance, and measurement error to generate 6,000 operating scenarios. Observable fault cases are selected using an explicit current-based criterion, after which a data-only multilayer perceptron, a physics-informed multilayer perceptron, Random Forest, Gradient Boosting, and HistGradientBoosting are evaluated on an independent test subset. A weighted ensemble, 0.4 MLP + 0.6 HGB, achieves a mean absolute error of 0.962 km, an RMSE of 1.316 km, and an R 2 of 0.948, reducing MAE by approximately 25% relative to the data-only MLP. Error distributions are further analyzed with respect to fault resistance, measurement noise, observability, and uncertainty. The proposed model is not intended to replace deterministic protection; instead, it operates as a confidence-gated advisory layer that narrows the patrol segment and abstains under unreliable conditions. An IEC 61850-compatible deployment architecture, cybersecurity controls, a TRL 3-7 validation roadmap, and the remaining requirements for HIL, COMTRADE, and seasonal field validation are also presented.

More from our Archive