Synchrophasor Driven Voltage Stability Assessment Using Adaptive Deep Learning Based Tools on Temporal Ensembling and Data Augmentation

This research explores the application of Phasor Measurement Units (PMUs) and deep learning methodologies to predict voltage stability by enhancing the efficient operation of the power grid. The study addresses the challenges of topological changes and contingency labeling in power system networks. The methodology incorporates generative artificial intelligence and semi-supervised learning by significantly improving the predictive accuracy of the various learning models. Least Squares Generative Adversarial Networks (LSGANs) strategically augment the training dataset, expanding both the feature space and temporal domain. The enriched dataset enables more accurate classification and enhances the model's reliability under previously unseen dynamic time-series signatures. The Recurrent Neural Network (RNN) based Temporal Ensembling improves the model's ability to determine voltage stability by clustering time signatures based on signal transitions and temporal dynamics. The deep learning model is applied to PMU time series data, which undergoes systematic evaluation criteria using various data preparation stages. In addition, the study explores multiple network topologies for the model's adaptability, testing across diverse time windows and signal conditions. Also, Hyperparameter tuning of the nominated model optimizes the performance through cross-validation, and the best configurations for the best settings were ranked based on the test scores. The findings underscore the potential of artificial intelligence and machine learning models to enhance power system stability. Such forecasts can support proactive decision-making, improve power system operations, and lay the foundation for future advancements in wide-area power system monitoring and control.