Imbalanced Sample Generation and Evaluation for Power System Transient Stability Using CTGAN

Gengshi Han, Shunyu Liu, Kaixuan Chen, Na Yu, Zunlei Feng, Mingli Song

Abstract

Power system Transient Stability Assessment (TSA) is critical for grid reliability but suffers from severe data imbalance, as unstable states are rare. This paper proposes a controllable data generation framework using Conditional Tabular Generative Adversarial Networks (CTGAN) to synthesize high-quality unstable samples. By employing Gaussian Mixture Model (GMM) normalization and multi-condition vector transformation, the framework generates realistic tabular data that preserves complex grid behaviors. Evaluation on the IEEE 39-bus system demonstrates that augmenting training sets with CTGAN-generated samples improves the recall of unstable cases in deep learning models, enhancing the accuracy of real-time stability predictions.

Introduction

Reliable power grid operation depends on Transient Stability Assessment (TSA)—the ability of a system to maintain synchronism after a major disturbance like a short circuit. While traditional Time Domain Simulations (TDS) are highly accurate, their computational intensity makes them unsuitable for real-time decision-making. Deep learning (DL) models offer a promising, high-speed alternative. However, DL models require large, balanced datasets to perform effectively.

In practice, power systems operate in stable states the vast majority of the time. This leads to a "data scarcity" problem for unstable samples. When trained on such imbalanced data, DL models develop a bias toward predicting stability, potentially missing rare but catastrophic failure modes. This research addresses the imbalance by introducing a CTGAN-based framework to generate synthetic, controllable, and statistically accurate unstable samples.

Methodology

The proposed framework leverages CTGAN, a generative model specifically optimized for tabular data. Unlike standard GANs, CTGAN is designed to handle the non-Gaussian and multi-modal distributions typical of power system variables like bus voltages and generator angles.

Overview of the CTGAN-based sample generation process, including GMM normalization and conditional vector transformation.

The methodology consists of three primary stages:

  1. Data Preprocessing with GMM: Instead of standard linear scaling, the authors use a Gaussian Mixture Model (GMM) to normalize tabular data. This preserves the specific "modes" or clusters of operating conditions, ensuring the generator learns the true underlying physics of the grid.
  2. Conditional Vector Design: To enable "controllable" generation, multiple conditions (e.g., stability labels and load levels) are encoded into a single one-hot vector. This allows the user to specify the characteristics of the generated samples, such as requesting an "unstable" sample at a "110% load level."
  3. Adversarial Training: The CTGAN architecture uses a generator and a discriminator with fully connected layers. The generator learns to produce samples that the discriminator cannot distinguish from real simulation data, while the conditional input ensures the output matches the desired state.

Experimental Setup

The framework was tested using the IEEE 39-bus system, a standard benchmark representing the New England power grid. Initial "real" data was produced via Time Domain Simulations using Matpower and PSAT. The resulting dataset contained 14,221 samples with a roughly 4:1 ratio of stable to unstable cases. Each sample included critical features such as voltage magnitudes, angles, and active/reactive power at the moment of fault clearance.

Evaluation Framework

The authors emphasize that synthetic data must be evaluated for more than just visual similarity. They propose a three-pronged evaluation strategy to ensure the generated samples are both realistic and useful.

The three-part evaluation framework: testing conditional control, statistical distance, and downstream model utility.
  1. Conditional Generation: Verifying if the model produces the requested class (e.g., stable vs. unstable). Results showed that when conditioned for instability, the proportion of unstable cases rose from ~40% to nearly 60%.
  2. Distance Calculation: Using JS Divergence and Wasserstein distance to measure how closely the synthetic distribution matches the real distribution. These metrics confirmed that synthetic samples were nearly indistinguishable from real simulation data.
  3. Performance of Assessment Models: Testing if adding synthetic data to a training set improves the performance of downstream classifiers like Decision Trees and Multi-Layer Perceptrons (MLPs).

Key Results and Insights

The results demonstrate the effectiveness of the CTGAN approach in bridging the data gap. The most significant finding was the impact on downstream models. Adding CTGAN-generated samples to the training set increased the recall for unstable samples—the most critical metric for grid safety—by 1.48% for Decision Trees and 2.74% for MLPs. Overall F1-scores and accuracy also saw improvements, proving that the synthetic data provided meaningful information rather than just noise.

Conclusion and Future Work

This research provides a robust solution to the data imbalance problem in power system stability testing. By using CTGAN with GMM normalization, the framework generates high-quality synthetic data that improves the reliability of deep learning-based TSA.

However, the study focuses on a single test system (IEEE 39-bus). Future work will explore how this framework scales to much larger, real-world utility grids with thousands of buses. Additionally, researchers aim to explore more complex conditions, such as varying fault locations or different types of renewable energy penetration, to further enhance the model's robustness.

Summary updated

BibTeX
@inproceedings{han2021imbalanced,
  author = {Gengshi Han and Shunyu Liu and Kaixuan Chen and Na Yu and Zunlei Feng and Mingli Song},
  title = {{Imbalanced Sample Generation and Evaluation for Power System Transient Stability Using CTGAN}},
  booktitle = {ICO},
  year = {2021},
  url = {https://openreview.net/forum?id=32F35EpGHo},
}
Figure