Date of Award
2024
Degree Type
Dissertation
Degree Name
Doctor of Philosophy in Electrical Engineering
Department
Electrical, Computer, and Biomedical Engineering
First Advisor
Yan Sun
Second Advisor
Hui Lin
Abstract
The future of power generation and delivery relies on the safe and secure operation of the smart grid. Intelligent devices and networked automation and control make the smart grid cleaner and more efficient to meet society’s ever-changing energy needs. At the same time, the advent of the smart grid introduces new vulnerabilities to this critical infrastructure. The security and resiliency of the smart grid are paramount to national security and will require an approach that leverages traditional network security in addition to domain-specific anomaly detection.
Data-driven anomaly detection has become a prevalent research topic for smart grids in recent years, owing to the availability of data in the grid and advancements in computation. However, there exists a research gap in how to create meaningful training sets for data-driven methods that accurately represent the grid’s changing operating conditions. In this research, we propose physics-aware data augmentation to enhance fault detection.
The first proposed method uses a unique prediction-based data augmentation technique to create a meaningful training set that accurately represents the changing operating conditions of the grid. We leverage load prediction and smart grid simulations to predict future normal and faulty operations to incorporate with existing training data. In doing so, we enhance the detection and locating of bus faults and transmission line outages in IEEE benchmark systems using multiple data-driven approaches. The results demonstrate the benefits of prediction-based data augmentation become even more pronounced in instances where more historical data is unavailable.
Second, we study data augmentation under the conditions of a cyber-physical attack. A simple attack model is used to demonstrate that an attacker could significantly impact the performance of pre-trained data-driven anomaly detection by denying data in a small region that includes the line outage. We propose a novel conditioned neural network algorithm that provides a deep learning procedure to incorporate knowledge of that data availability during testing. Our conditioned neural network utilizes a pre-trained historical model as an input, in addition to training data, which has been conditioned with the knowledge of current conditions in the grid, to return the functionality of deep learning classification for the data denied region. Furthermore, we study this algorithm in combination with the temporal data augmentation gained through prediction.
The results in this dissertation demonstrate the significant impact of this research on the generalization of deep learning to the power grid. Prediction-based data augmentation improves the detection of bus faults by 8.6% on average, and line outage detection improves by 4.19%. Spatial data augmentation through conditioning improves the detection of line outages in a data-denied environment by 56%. The success of deep learning algorithms depends greatly on the composition of the training data and its ability to generalize well to the testing conditions. In the smart grid, significant amounts of data are needed to capture all the dependencies and variables. A physics-informed data augmentation method can improve the detection of faults without increasing the size of the data or retraining the historical model.
Recommended Citation
Rogers, Jennifer, "DATA AUGMENTATION FOR SMART GRID ANOMALY DETECTION" (2024). Open Access Dissertations. Paper 1659.
https://digitalcommons.uri.edu/oa_diss/1659