A Generative AI-enhanced framework for intrusion detection and resilient Transient Stability Prediction in Networked Cyber–Physical Power Systems

Networked Cyber-Physical Power Systems (NCPPSs) rely on wide-area communication and Phasor Measurement Unit (PMU) streams, which makes data-driven Transient Stability Prediction (TSP) vulnerable to stealthy False Data Injection Attacks (FDIAs). This paper proposes a latency-aware intrusion detection, recovery, and resilient TSP framework for NCPPSs. The framework combines a dual Weighted Least Squares (WLS)/Square-Root Unscented Kalman Filter (SR-UKF) estimation layer, a hybrid Convolutional Neural Network-Long Short-Term Memory (CNN–LSTM) Intrusion Detection System (IDS), a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) adversarial autoencoder for active signal restoration, and a Deep Transformer ensemble classifier. Unlike IDS-only, estimation-only, or recovery-only baselines, the proposed framework links detection, generative restoration, estimator fallback, and stability prediction in a single resilience pipeline. Tests on the Northeast Power Coordinating Council (NPCC) 48-machine benchmark show that WGAN-GP restoration reduces rotor-speed root mean square error by about 82% and bus-voltage-magnitude mean absolute error by more than 93%, allowing the TSP classifier to maintain about 99% accuracy when 50% of PMU streams are compromised. The design shifts TSP security from passive alarm generation to active prediction continuity under cyber–physical disturbances. Prototype benchmarking gives 233 ms total latency, with only 18 ms from the IDS and recovery modules, indicating practical feasibility for optimized edge-assisted deployment.