Funded by NSF (Co PI, 200K, Jun/2024 – May/2027)
Global warming not only causes rapid climate change and destroys our ecosystem but also increases global economic and social inequality. In response, nations worldwide are striving to integrate renewable energy sources into the existing power grid to reduce fossil-fuel consumption. As the energy resource diversifies and sensing and control increases, the power grid grows into a smart grid. Compared with the traditional fuel-based power grid that only has uni-source power connections, the multi-source power grid suffers from scalability, stability, and multi-modality issues caused by many heterogeneous power generators.
An accurate, real-time, and scalable anomaly detection and diagnosis method that can protect the system is highly desirable yet extremely challenging. Due to the reliance on information and communication technologies, the smart grid faces a significant risk in security. For example, the attacker’s malicious modification of controllers or sensors in power electronics converters, such as in PV farms and wind farms, would affect both power electronics and the grid, leading to catastrophic failures and substantial economic losses. However, existing anomaly detection methods suffer from inherited limitations in one or more of the following aspects: sensitivity, specificity, scalability, stability, multi-modality, and computational complexity.
Considering the uniqueness of smart grids that include many power electronics systems, developing data-driven algorithms that can adapt to changing conditions and network topology is more challenging. The investigators share a synergistic vision and passion and have been collaborating for extensive periods. In this project, investigators generalize their recent development on graph-on-graphs methods to a family of methods for smart grid anomaly detection and diagnosis, considering the dynamic association between multi-modality data. Unlike existing methods that have degraded performance for data with spurious noise, the proposed methods are highly sensitive and precise. More importantly, the proposed real-time methods are computationally efficient. In general, the graph-on-graphs approach provides a rich and flexible framework to address anomaly detection in multi-modality networks. The results from this project will make significant theoretical and methodological contributions to anomaly detection and diagnosis in smart grids and have a strong impact on smart grid developments. This project crosses the boundaries between statistics, computer science, and engineering and will provide a unique opportunity to attract students of different scientific, ethnic, and national backgrounds to study science and engineering to foster our next-generation scientists and engineers.