World Journal of Engineering Research and Technology

Nsikak E. Udoh*, Nseobong I. Okpura, Kingsley M. Udofia

ABSTRACT

Weak distribution grids, such as the Nigerian 33‑kV network, suffer from reactive fault management, leading to prolonged outages and poor reliability. This paper developed and validated a dual‑layer bagged decision tree ensemble for real‑time fault state and type classification using three years (2022–2024) of hourly SCADA data from a primary feeder. A feature engineering algorithm extracts lag variables, rate‑of‑change metrics, rolling statistics, and frequency deviation. Layer 1 classifies the system state as Normal, Fault, or Maintenance, while Layer 2 further categorises fault instances into EarthFault, OverCurrent, or PowerFailure. Using temporal forward chaining (training on 2022–2023, testing on 2024), the ensemble achieved a Layer 1 accuracy of 84.2% and a Layer 2 accuracy of 88.4%. Area under the ROC curve (AUC) reached 0.98 for PowerFailure and 0.95 for Maintenance, with inference latency of 64 ms per sample — well within typical SCADA polling intervals. Feature importance identified historical power trend and current rate of change as the most predictive variables. While the model reduced false alarms compared to conventional overcurrent relays, 1,562 false positives (Normal as Fault) remained an operational concern and can be mitigated through probability threshold tuning. The framework provides a computationally lightweight, interpretable solution for near‑real‑time fault awareness in resource‑constrained environments, though infrastructural hardening remains essential. Future work will explore online learning and transfer to other feeders.

[Full Text Article] [Download Certificate] https://doi.org/10.5281/zenodo.21159531