Real-Time Machine Learning Models To Detect Cyber And Physical Anomalies In Power Systems

A Smart Grid is a cyber-physical system (CPS) that tightly integrates computation and networking with physical processes to provide reliable two-way communication between electricity companies and customers. However, the grid availability and integrity are constantly threatened by both physical faults and cyber-attacks which may have a detrimental socio-economic impact. The frequency of the faults and attacks is increasing every year due to the extreme weather events and strong reliance on the open internet architecture that is vulnerable to cyber-attacks. In May 2021, for instance, Colonial Pipeline, one of the largest pipeline operators in the U.S., transports refined gasoline and jet fuel from Texas up the East Coast to New York was forced to shut down after being attacked by ransomware, causing prices to rise at gasoline pumps across the country. Enhancing situational awareness within the grid can alleviate these risks and avoid their adverse consequences. As part of this process, the phasor measurement units (PMU) are among the suitable assets since they collect time-synchronized measurements of grid status (30-120 samples/s), enabling the operators to react rapidly to potential anomalies. However, it is still challenging to process and analyze the open-ended source of PMU data as there are more than 2500 PMU distributed across the U.S. and Canada, where each of which generates more than 1.5 TB/month of streamed data. Further, the offline machine learning algorithms cannot be used in this scenario, as they require loading and scanning the entire dataset before processing. The ultimate objective of this dissertation is to develop early detection of cyber and physical anomalies in a real-time streaming environment setting by mining multi-variate large-scale synchrophasor data. To accomplish this objective, we start by investigating the cyber and physical anomalies, analyzing their impact, and critically reviewing the current detection approaches. Then, multiple machine learning models were designed to identify physical and cyber anomalies; the first one is an artificial neural network-based approach for detecting the False Data Injection (FDI) attack. This attack was specifically selected as it poses a serious risk to the integrity and availability of the grid; Secondly, we extend this approach by developing a Random Forest Regressor-based model which not only detects anomalies, but also identifies their location and duration; Lastly, we develop a real-time hoeffding tree-based model for detecting anomalies in steaming networks, and explicitly handling concept drifts. These models have been tested and the experimental results confirmed their superiority over the state-of-the-art models in terms of detection accuracy, false-positive rate, and processing time, making them potential candidates for strengthening the grid's security.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.