A novel framework for real time transient stability prediction in power systems under data integrity attacks

Abstract

This thesis presents a comprehensive and innovative framework for addressing one of the most critical challenges in modern power systems—maintaining transient stability under the threat of cyber attacks, particularly false data injection attacks (FDIAs). The work integrates robust dynamic state estimation with state-of-the-art machine learning algorithms and dedicated cyber-resilient measures to ensure reliable and accurate real-time transient stability prediction even as power grids grow in complexity. The framework is organized into two primary components. The first component focuses on high-fidelity data acquisition from Phasor Measurement Units (PMUs), which deliver high-frequency, synchronized electrical measurements. These measurements are enhanced by dynamic state estimates obtained via a Square-Root Unscented Kalman Filter (SR-UKF). By incorporating dynamic state variables such as rotor angles, rotor speeds, and internal voltage components, the framework creates an enriched feature space that significantly improves the fidelity of transient stability predictions. To mitigate cybersecurity risks, the thesis introduces an additional cyber-resilient layer. This layer features a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) anomaly detector designed to identify subtle deviations in estimation residuals indicative of FDIAs. Upon detection of a potential cyber attack, a measurement recovery module—based on a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) adversarial autoencoder (AAE)—is invoked to reconstruct and restore compromised data prior to stability analysis. The predictive aspect of the framework is driven by two ensemble approaches. One approach is a voting classifier that combines several machine learning models—including a Multi-Layer Perceptron (MLP), Random Forest (RF), Support Vector Machine (SVM), Convolutional Neural Network (CNN), and Light Gradient Boosting Machine (LightGBM)—using a weighted soft voting mechanism. This method effectively leverages the complementary strengths of the individual models to achieve robust transient stability predictions under noisy and adversarial conditions. The second approach is a stacking ensemble that integrates predictions from deep transformer base learners through an XGBoost meta-learner, providing an additional layer of refinement to the overall prediction accuracy. Extensive simulations on the benchmark NPCC 48-machine, 140-bus system demonstrate significant improvements in prediction accuracy when dynamic state estimates are incorporated. The results convincingly show that the integrated framework not only enhances prediction performance but also robustly mitigates the adverse effects of FDIA, rendering it highly suitable for real-time operational deployment. From an academic standpoint, this thesis makes a notable contribution by addressing the dual challenges of transient stability prediction and cyber resilience. It pushes the frontier of research by combining model-based dynamic state estimation with advanced machine learning techniques to improve both accuracy and reliability. Future work might focus on scaling the framework for larger systems, further robustness evaluations under different noise and attack scenarios, and integration into existing control center architectures. Overall, the proposed methodology represents a significant advancement in the field of power system stability analysis.