Multi-class imbalanced learning for short-term voltage stability assessment

Abstract

Imbalanced databases tend to bias machine learning models toward the majority class, compromising the accuracy of network state assessment and leading to suboptimal or erroneous decision-making. This study addresses the issue of data imbalance by proposing a synthetic data generation approach based on a Generative Adversarial Network (GAN). The proposed model employs a conditional Wasserstein GAN with a gradient penalty. A Gated Recurrent Unit (GRU) network integrated with an attention mechanism is utilized to generate diverse, high-quality, and realistic data. The experiments are conducted on the IEEE 118-bus and a real-world network. The findings show that the proposed method can effectively produce realistic, high-quality samples for minority classes. In addition to accuracy, performance is evaluated using metrics such as Misdetection (Mis), False Alarm (FA), and G-mean. The model’s robustness is validated under topology changes and varying imbalance ratios. Findings from the real-world network demonstrate resilient performance and promising results in STVS assessment.