Physics Learning Based Stealthy False Data Attack Against Renewable Fed Uncertain Power System

Despite high success rates, the performance of existing model-free false data attack (MFFDA) methods against state estimation (SE) in power systems with huge renewable penetration is questionable. They lack sparsity and accountability for generator and zero bus injection, which can be detected by control centres. Increased uncertainty from renewable energy sources further heightens detection risks. This paper assesses SE security against an advanced MFFDA. We propose a hybrid physics-deep learning Autoencoder—Generative Adversarial Network (AE-GAN) framework to capture the inherent variability in measurement data and generate attack data. The AE-based proxy SE model incorporates the temporal complexities of uncertain measurements. We also present a refined training data algorithm, utilizing generator and renewable power equivalences in a sparse attack area to enhance stealth. Validation on the IEEE 14 and 118 bus systems and analytical studies indicate that our method is effective with a high success rate.