GAN-based SSO Synthetic Data Generation for Improving the Performance of Deep Learning Classification Models

Gathering real-world oscillation data is challenging, even though the penetration of inverter-based resources (IBRs) contributes to sub-synchronous oscillations (SSO) in power systems. This paper proposes a synthetic data-generation process based on Deep Convolutional Generative Adversarial Network (DCGAN) to generate data and enhance the training performance of deep learning models. To ensure the synthetic data mimics the actual data patterns, Prony analysis was employed to compare phase and magnitude characteristics. Simulation results demonstrated that DCGAN-generated waveforms achieved an average 92.3% similarity to real data in Prony-based evaluations. To further improve generation quality, a clustering approach was applied, grouping similar-patterned data before training, which resulted in more realistic oscillation patterns. The impact of synthetic data augmentation was evaluated using K-Nearest Neighbors (KNN) for SSO anomaly detection. Data augmented with synthetic data yielded improved classification results, with accuracy increasing by up to 0.2% and F1 score improving by 0.1% on average. Additionally, compared to traditional Simulink-based simulation, DCGAN was 46.25 times faster in generating synthetic oscillation data, significantly reducing the computational cost of obtaining large datasets.