When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.
Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.
The evolution of the electric power grid into a fully digitized, automated cyber-physical system is critical to optimizing demand response and ensuring a sustainable future. However, the smart grid's growing dependency on digital technologies and networks, ranging from advanced metering systems to distributed control operations, unveils novel security and resilience challenges. The resilience of the grid can be compromised by both malicious attacks and natural disasters, which can target either the physical infrastructure or cyber systems, leading to disruptions or performance degradation. Considering its central role in the energy infrastructure, it's imperative to enhance its resilience and fortify it against possible cyber threats and natural disasters. The main goal of this thesis was to advance machine learning-based models in order to enhance the security and resilience of the smart grid. This objective was realized through a detailed examination of notable cyberattack to identify the most detrimental scenarios, thereby aiding in the development of superior countermeasures, and the establishment of a predictive model to anticipate service degradation, thus improving service reliability. Due to the potentially disastrous repercussions in terms of resilience by misguiding operators with undetected manipulation of measurements, as a cyberattack use case, the research initially focused on a form of data integrity attack, called false data injection attack (FDIA). The primary research focuses on analyzing the worst-case scenario by proposing a scheme that minimizes the attacker's knowledge and resources by reducing the number of attacked measurements without requiring knowledge of the grid's topology. By incorporating techniques such as principal component analysis and particle swarm optimization, the model effectively generates attack vectors with minimized norm, ensuring their feasibility and impact. The results demonstrate the scheme's capability to generate undetectable attack vectors while maintaining significance, even under restricted topology knowledge and compromised meters. In addition, a multi-objective optimization approach was explored to analyze FDIA in power system state estimation. Through the use of the Improved Strength Pareto Evolutionary Algorithm (SPEA2), the proposed approach efficiently minimizes the number of attacked meters while maximizing the impact of the attack. The findings reveal the possibility of implementing a stealthy and sparse FDIA, allowing attackers to optimize the expected impact on state vectors and branch power flows. Moreover, the impact analysis highlights the trade-off between stealthiness and impact, showing that larger systems can be significantly affected even with fewer compromised meters. Furthermore, the thesis addresses the potential exploitation of artificial intelligence (AI) by adversaries by proposing a blind FDIA scheme that relies solely on historical measurements for grid topology inference and attack vector generation. Inspired by black-box attacks, the model incorporates a substitute bad data detector to filter the attack vector, reducing the chance of detection. The evaluation framework validates the stealthiness and impact of this blind FDIA scheme, demonstrating the effectiveness of adversarial AI in refining the attack model and posing new challenges for security research in the smart grid and other cyber-physical systems. Lastly, a sophisticated prediction model was developed to anticipate service degradation in 5G communication-based distributed feeder automation systems. This model integrates a mutual information based FastICA as an anomaly indicator providing dissimilarity indices between normal and abnormal data, a dual discriminator conditional generative adversarial network (cGAN) for abnormal data prediction, and fuzzy logic for classification. By training on three distinct scenarios, the model effectively predicts future abnormal scenarios based on current normal scenarios, acting as an early warning system for potential service degradation and facilitating preventative measures. This contributes to the improvement of service reliability and enhances the overall security and resilience of these networks. Overall, this thesis endeavors to make significant contributions to enhancing the security and resilience of smart grids by addressing worst-case cyberattack investigations revealing AI's potential in the cyberattack plane, and providing innovative machine-learning strategies for predicting service degradation, ultimately serving to improve countermeasures and prevention mechanisms to fortify the resilience and reliability of these critical cyber-physical systems.