Abstract
The transition toward low-inertia power systems, driven by the integration of non-synchronous renewable energy, necessitates rapid and automated event diagnosis. This paper presents a framework for power system event classification (PSEC) using wide-area Phasor Measurement Unit (PMU) data. By integrating Linear Discriminant Analysis (LDA) for dimensionality reduction and Sequential Forward Selection (SFS) for feature optimization, the proposed method identifies a minimal set of critical features. A Quadratic Discriminant Analysis (QDA) classifier then achieves near-perfect accuracy with significantly reduced computational overhead. The approach is validated using the IEEE-39 Bus Test System, demonstrating a 10x speedup in training time compared to raw data processing.
Introduction
Modern electrical grids are undergoing a fundamental shift as traditional synchronous generators are replaced by intermittent, non-synchronous renewable sources like wind and solar. This transition creates "low inertia" systems that are far more sensitive to sudden changes, meaning small faults can escalate into large-scale blackouts much faster than before. While Wide-Area Monitoring Systems (WAMS) using PMUs provide high-resolution data, the sheer volume is overwhelming for manual processing. There is a critical need for automated systems that can instantly diagnose events—such as generation dips or line trips—to support real-time decision-making.
Methodology
The proposed PSEC framework utilizes a multi-stage pipeline designed for both speed and accuracy.
- Feature Extraction (LDA): Raw PMU data is high-dimensional (261 features). LDA projects this data into a lower-dimensional subspace while maximizing the separability between different event types.
- Feature Selection (SFS): To further optimize, SFS—a "greedy" algorithm—selects only the features that significantly reduce cross-validation error.
- Classification (QDA): QDA is employed for final classification as it effectively handles the non-linear, overlapping boundaries inherent in complex power system dynamics.
- Adaptive Retraining: The system is designed to be self-training, incorporating successfully classified new events back into the database to improve performance over time.
Experimental Setup and Datasets
The study utilized a dynamically simulated dataset based on the IEEE-39 Bus Test System. The dataset included 397 events across six categories: Generation Dip (GD), Loss of Load (LL), Line Trip (LT), Capacitor Switching, and Synchronous Motor Start. Data was captured from 29 PMUs sampling at 100 Hz, recording frequency, phase angle, and voltage measurements.
Results and Discussion
A key finding was that out of 261 potential features, the system required only eight specific features from five buses to maintain high accuracy. This demonstrates that a highly observant monitoring system does not require a PMU on every bus.
The combination of LDA and SFS significantly improved computational efficiency. Training time dropped from 32.8 seconds (raw data) to approximately 3.06 seconds—a 10x speedup essential for real-time applications. The QDA classifier achieved 100% accuracy on the test set using the reduced 8-feature set, outperforming Support Vector Machines (SVM) and k-Nearest Neighbor (k-NN) algorithms in both speed and accuracy.
Conclusion
This research demonstrates that automated, high-speed event classification is achievable in low-inertia grids by focusing on a small subset of highly informative features. Future work will focus on incorporating real-world PMU data noise and expanding the model to include more complex scenarios like transformer energization and battery storage cycles.
Summary updated