Blending data and model for robust and secure power system operation
File(s)
Author(s)
Xu, Wangkun
Type
Thesis
Abstract
The evolving landscape of power systems, characterized by the trend of decarbonization, digitalization and decentralization, demands more efficient, robust, and secure operation strategies. Traditional model-based approaches are being challenged, leading to a transition to data-driven methods enabled by advances in information and communication technologies. However, concerns persist regarding the interpretability and reliability of purely data-driven decision-making processes. Hence, this thesis explores an intermediate approach that blends the data and the model for power system operation, offering a viable solution to the new challenges.
Two distinct frameworks are examined, each offering varying degrees of integration. The first framework orchestrates sequential learning and optimization processes to facilitate the exchange of critical information. The second framework embeds optimization models within deep learning structures, enabling the forecast to be decision-aware.
Chapter 2 presents a robust moving target defence method for the detection of false data injection attacks. By optimizing the set points of distributed flexible AC transmission system devices in real-time, the method maximizes the detection probability under specific measurement noise levels. Within the context of sequential design in Chapter 3, the thesis illustrates how a data driven attack detector and physics-informed attack identifier can spatially and temporally reduce the operational cost of robust moving target defence by quantifying its uncertainty set. The sequential design instills greater trust among system operators, compared to its pure data-driven counterpart.
Chapter 4 evaluates the generalizability of the integrated framework. A unified adversarial training approach is proposed to address its uncertainties in both the input space of the deep neural network and the parameter spaces of model-based optimization. In Chapter 5, the integrated framework is introduced to facilitate machine unlearning tasks in load forecasting, providing a balance between data privacy and the operation cost of the whole system.
Two distinct frameworks are examined, each offering varying degrees of integration. The first framework orchestrates sequential learning and optimization processes to facilitate the exchange of critical information. The second framework embeds optimization models within deep learning structures, enabling the forecast to be decision-aware.
Chapter 2 presents a robust moving target defence method for the detection of false data injection attacks. By optimizing the set points of distributed flexible AC transmission system devices in real-time, the method maximizes the detection probability under specific measurement noise levels. Within the context of sequential design in Chapter 3, the thesis illustrates how a data driven attack detector and physics-informed attack identifier can spatially and temporally reduce the operational cost of robust moving target defence by quantifying its uncertainty set. The sequential design instills greater trust among system operators, compared to its pure data-driven counterpart.
Chapter 4 evaluates the generalizability of the integrated framework. A unified adversarial training approach is proposed to address its uncertainties in both the input space of the deep neural network and the parameter spaces of model-based optimization. In Chapter 5, the integrated framework is introduced to facilitate machine unlearning tasks in load forecasting, providing a balance between data privacy and the operation cost of the whole system.
Version
Open Access
Date Issued
2024-03
Date Awarded
2024-05
Copyright Statement
Creative Commons Attribution NonCommercial Licence
License URL
Advisor
Teng, Fei
Jaimoukha, Imad
Sponsor
Imperial College London
Publisher Department
Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)