Abstract
A two-stage machine learning-based approach for creating synthetic phasor measurement unit (PMU) data is proposed in this article. This approach leverages generative adversarial networks (GAN) in data generation and incorporates neural ordinary differential equation (Neural ODE) to guarantee underlying physical meaning. We utilize this approach to synthetically create massive eventful PMU data, which would otherwise be difficult to obtain from the real world due to the critical energy infrastructure information (CEII) protection. To illustrate the utility of such synthetic data for subsequent data-driven methods, we specifically demonstrate the application of using synthetic PMU data for event classification by scaling up the real data set. The addition of the synthetic PMU data to a small set of real PMU data is shown to have improved the event classification accuracy by 2 to 5 percent.
Introduction
Modern power grids rely heavily on Phasor Measurement Units (PMUs) to monitor system health and classify critical events such as faults or load shedding. However, the development of robust data-driven classifiers is significantly hindered by the scarcity of "eventful" data. Major grid disturbances are rare, and existing data is often restricted due to Critical Energy Infrastructure Information (CEII) security regulations.
While standard Generative Adversarial Networks (GANs) are proficient at generating high-dimensional data like images, they often struggle to capture the strict physical laws—specifically the underlying differential equations—that govern power system dynamics. This can result in synthetic data that appears physically impossible, potentially degrading the performance of downstream classifiers.
Methodology: A Two-Stage Hybrid Framework
The authors propose a novel two-stage generation approach that blends deep learning with classical physics to ensure both spatial and temporal fidelity.
Stage 1: Spatial Correlation via GAN
In the first stage, a vanilla GAN is trained to generate "snapshots" of grid events. This model captures the complex spatial relationships between different buses in the grid. As shown in Figure 1, the GAN architecture involves a Generative Model (G) that creates synthetic PMU data from random noise, while a Discriminative Model (D) learns to distinguish between real and synthetic samples through backpropagation.
Stage 2: Temporal Correlation via Neural ODE
To address the temporal limitations of standard GANs, the authors incorporate a Neural Ordinary Differential Equation (Neural ODE) model. Instead of forcing the GAN to "guess" the entire time series, the Neural ODE takes the initial state generated in Stage 1 and evolves it over time. Since power systems are physically defined by ODEs, this ensures the synthetic "tail" of the event follows realistic physical trajectories.
As illustrated in Figure 2, the hybrid framework uses Model G for the initial snapshots and Model f (the Neural ODE) to solve for the post-event time series. This combined approach ensures that the resulting synthetic eventful PMU data is both diverse and physically consistent.
Experimental Setup and Validation
The framework was tested using the IEEE 39-bus System, a standard benchmark for power system modeling. The study focused on three common event types: bus faults, line tripping, and load shedding. To simulate real-world data scarcity, the generative models were trained using only 20 real samples per event type.
To validate the quality of the synthetic data, the authors employed Prony analysis, a classical power systems tool used to extract oscillation frequencies and damping ratios. If the synthetic data’s "modes" match those of real data, it passes the fidelity test.
Key Results
- Accuracy Improvement: Across various classification pipelines (utilizing PCA, Wavelet Transforms, SVMs, and Ensemble Trees), augmenting small real-world datasets with synthetic data consistently improved event classification accuracy by 2% to 5%.
- High Fidelity: The synthetic data achieved an "Accuracy Test" score of 82% to 90% in modal analysis, indicating that the majority of synthetic samples were physically indistinguishable from real simulations.
- Diversity: The GAN-Neural ODE combination successfully generated new, meaningful variations of events rather than simply cloning the limited training samples, allowing classifiers to generalize better to unseen data.
Conclusion and Future Work
By "baking" physical requirements into the model architecture through Neural ODEs, this research provides a path forward for generating high-fidelity synthetic data in domains governed by strict physical laws. Future work will focus on scaling this approach to handle the high-dimensional state spaces of real-world utility grids involving thousands of buses and investigating automated event detection to identify the necessary timestamps for the generation process.
Summary updated