Fast diffusion-driven data augmentation for imbalanced power-system stability classification

The increasing integration of intermittent renewable energies amplifies uncertainty in power system operation, complicating stability assessment and limiting the effectiveness of traditional model-based approaches. To address this challenge, data-driven methods have emerged as flexible alternatives, but they often suffer from severe class imbalance, with unstable cases being relatively rare. Generative models like Generative Adversarial Models (GANs) and Variational Autoencoders (VAEs) have been explored for sample augmentation, yet frequently encounter issues such as mode collapse and insufficient diversity. More recently, Denoising diffusion probabilistic models (DDPMs) have shown promise, but their high computational cost—often requiring hundreds of denoising steps—hinders practical deployment. To overcome these limitations, this paper proposes an FDDA that directly learns the residuals between noise samples at adjacent time steps, thereby eliminating the need for larger attention mechanism modules. This design significantly reduces the number of diffusion steps by more than 60% compared to standard DDPMs—without compromising sample quality. Combined with support vector machine(SVM) and random forest(RF) classifiers, our approach is evaluated on multiple benchmark power system cases. Results demonstrate improved generalization, robustness to class imbalance, and lower computational cost compared to conventional augmentation techniques, thus providing a scalable and efficient data-driven strategy for enhancing power system stability assessment under high uncertainty. • Diffusion-driven augmentation reduces class imbalance in power-system stability data. • Compared with baselines, improves fault-scenario diversity and sample quality. • Combining support vector machines and random forests improves assessment accuracy.