Review of Low-Rank Data-Driven Methods Applied to Synchrophasor Measurement

Meng Wang, Joe H. Chow, Denis Osipov, Stavros Konstantinopoulos, Shuai Zhang, Evangelos Farantatos, Mahendra PatelIEEE Open Access Journal of Power and Energy 2021 · DOI

Abstract

There is a growing acceptance of using synchrophasor data collected over large power systems in control centers to enhance the reliability of power system operations. The spatial and temporal nature of power system ambient and disturbance response allows the analysis of large amount of synchrophasor data by low-rank methods. This paper provides an overview of several applications of synchrophasor data utilizing the low-rank property. The tools to capitalize on the low-rank property include matrix completion methods, tensor analysis, adaptive filtering, and machine learning. The applications include missing data recovery, bad data correction, and disturbance recognition.

Introduction: The Challenge of Grid Visibility

Phasor Measurement Units (PMUs) serve as the high-resolution "eyes" of the modern power grid. By providing real-time data on voltage and current phasors, they are essential for maintaining stability. However, the utility of this data is frequently compromised by quality issues: communication delays lead to data gaps, hardware malfunctions cause sensor errors, and cyber-attacks can inject malicious "bad data" to deceive operators.

Traditional approaches to data correction often require a precise physical model of the grid's topology and electrical parameters. This paper reviews emerging data-driven methods that bypass the need for complex physics models by exploiting the mathematical structure of the data itself—specifically, its "low-rank" property.

The Low-Rank Nature of Synchrophasor Data

Power grid data is not random noise; it is highly structured. Because generators across a region often move in synchrony and nearby buses experience similar voltage fluctuations, a large matrix of PMU measurements can be described by a small number of "dominant modes."

Synchrophasor measurements showing magnitude and phasor angle over time during a grid event.

As shown in the figure above, the temporal patterns (magnitude and angle) across different channels exhibit significant correlation. This spatial and temporal redundancy is what allows researchers to treat grid data as low-rank matrices or tensors, enabling the recovery of missing information even when significant portions of the data are lost.

Methodological Overview

The authors categorize several low-rank techniques designed to handle specific data quality challenges:

  • Matrix Completion & Hankel Matrices: While standard algorithms like Singular Value Thresholding (SVT) work for random data loss, they fail during "bursty" communication outages. By stacking data into a Hankel Matrix, temporal correlations are preserved, allowing algorithms like AM-FIHT to recover data even when all channels fail simultaneously.
  • Tensor Decomposition: Organizing data into 3D tensors (e.g., Channel × Time × Measurement Type) preserves multi-dimensional relationships that 2D matrices might flatten.
  • Robust PCA & Structured Alternating Projections (SAP): These methods are used to isolate "bad data" (outliers). SAP is particularly effective at correcting simultaneous errors across multiple channels, a scenario where traditional Robust Principal Component Analysis often struggles.
  • Koopman Theory: For severe, nonlinear disturbances, "lifting" techniques map data into a higher-dimensional space where it regains a linear, low-rank structure suitable for analysis.
Approximation error as a function of the approximate rank r for different Hankel matrix parameters κ.

Key Insights and Results

1. Superior Data Recovery

The AM-FIHT algorithm significantly outperformed traditional methods in scenarios where 50% of data was missing across all channels. The use of Hankel structures allows the model to "see" through consecutive gaps that would otherwise appear as empty columns in a standard matrix.

2. Enhanced Disturbance Classification

When training Convolutional Neural Networks (CNNs) to identify grid faults (e.g., generator trips or short circuits), training on the "low-rank subspace"—the underlying modes of the data—proved more effective than using raw time-series data. A feature-based CNN (CNN-F) achieved an Identification Accuracy Rate roughly 15% higher than standard models and showed greater resilience when training data was scarce.

3. Model-Less Resilience

A primary advantage of these methods is their ability to function without an up-to-date map of the grid. This makes them highly resilient to unexpected topology changes or situations where the physical parameters of the network are unknown.

Limitations and Future Directions

While low-rank methods are powerful, several hurdles remain for wide-scale deployment:

  • Uncertainty: Current methods provide a "best guess" for missing data but lack an uncertainty index to inform operators of the reliability of that guess.
  • Scalability: Processing data from thousands of PMUs in real-time requires further optimization of computational complexity.
  • Integration: Future work will likely focus on using low-rank methods as a pre-processing step to simplify complex PMU data for more advanced machine learning applications.

Summary updated

Abstract

There is a growing acceptance of using synchrophasor data collected over large power systems in control centers to enhance the reliability of power system operations. The spatial and temporal nature of power system ambient and disturbance response allows the analysis of large amount of synchrophasor data by low-rank methods. This paper provides an overview of several applications of synchrophasor data utilizing the low-rank property. The tools to capitalize on the low-rank property include matrix completion methods, tensor analysis, adaptive filtering, and machine learning. The applications include missing data recovery, bad data correction, and disturbance recognition.
BibTeX
@article{wang2021review,
  author = {Meng Wang and Joe H. Chow and Denis Osipov and Stavros Konstantinopoulos and Shuai Zhang and Evangelos Farantatos and Mahendra Patel},
  title = {{Review of Low-Rank Data-Driven Methods Applied to Synchrophasor Measurement}},
  journal = {IEEE Open Access Journal of Power and Energy},
  year = {2021},
  doi = {10.1109/oajpe.2021.3090579},
}
Figure