On Energy Theft Attack Detection in Smart Grids Using Machine Learning

Althobaiti, Ahlam and Marnerides, Angelos and Rotsos, Charalampos (2023) On Energy Theft Attack Detection in Smart Grids Using Machine Learning. PhD thesis, Lancaster University.

Text (2023AhlamAlthobaitiPhD)
2023AhlamAlthobaitiPhD.pdf - Published Version
Download (6MB)

Abstract

The convergence of legacy power system components with advanced information and communication facilities has led to the emergence of smart grids. Smart grids are envisioned to be the next generation of innovative power systems, guaranteeing resilience, reliability and sustainability, and facilitating energy production, distribution and management. Nonetheless, the development of such systems entails challenges covering a broad spectrum, ranging from operational management to data-driven power accounting and network security. Given the highly distributed properties of the modern grid, energy theft attacks can now be observed at various transmission and distribution levels. Apart from the financial gains for malicious actors, energy theft can also affect critical grid processes and have a direct impact on the grid’s overall resilience and safety. Conventional energy theft detection approaches rely on physically inspections, which are time-consuming, inaccurate, costly and require substantial human labour. By virtue of the smart grid paradigm, these inspections are now conducted more efficiently using modern data-driven and machine learning based detection approaches. Therefore, the major focus of this thesis is on designing a data-driven energy theft detection framework, taking advantage of the unique characteristics of modern smart grids. In particular, this thesis investigates and surveys the advances in energy theft strategies, as well as detection methods, from different perspectives on the smart grid, revolving around energy data manipulation of all three functions of demand, supply and generation. In addition, this thesis proposes a supervisory control and data acquisition (SCADA)-agnostic power modelling scheme for distributed renewable energy sources (DRES). Through this study, it is demonstrated that a viable and exogenous profiling solution achieving similar accuracy to SCADA-based schemes but under much lower computational time is required to produce a reliable regression model for DRES generation energy. Building on this work on SCADA-agnostic DRES power modelling, this thesis also describes a predictive energy theft detection approach for DRES. Evidently, the proposed approach yields a high DRES-based energy theft detection accuracy rate of over 95%, with low computational time required to produce DRES theft classification. Thus, it reasonably addresses the highly demanding requirements of low-cost and accurate real-time energy theft detection in modern power grids. Finally, this thesis introduces a self-learning theft detection system capable of distinguishing the properties of power consumption and generation theft with possible misconfigurations caused by nonmalicious intent. The proposed approach is adaptive through a self-learning operation that is continuously updated as new measurements become available. The results obtained indicate that this scheme can achieve over 90% accuracy in identifying theft with optimal over-streamed data measurements. Thus, it offers low computational time being required to classify consumption and generation meters, and its properties can be exploited for next-generation cross-batch energy theft detection.