Date of Award
May 2026
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Electrical Engineering
First Advisor
Prakash Ranganathan
Abstract
The rapid integration of distributed energy resources (DERs) into the modern power grid has shifted the energy landscape toward decentralized, smarter architectures. While beneficial for resilience and efficiency, this transition significantly expands the cyber attack surface of distributed energy resource management systems (DERMS). This dissertation addresses the vulnerabilities in the attack surface by proposing a comprehensive security framework and developing advanced, unsupervised machine learning and deep learning models for attack detection.
A Hierarchical Hybrid DERMS Federated Learning Framework and Intrusion Diagnostic Federated Framework are proposed, mapping grid entities across five levels to establish robust and scalable cross-layer security requirements. The research empirically validates detection of attacks at the sensing layer through investigations into phasor measurement units (PMUs), demonstrating that unsupervised clustering methods i.e., Density-Based Spatial Clustering of Applications with Noise (DBSCAN) and simple moving average (SMA) can detect false data injection (FDI) attacks with an accuracy of up to 98.4%. Additionally, grid visibility is maintained using the Amelia II method for data imputation, which achieved a mean absolute percentage error (MAPE) of 0.60 for reconstructing missing current data.
The core of this research leverages the non-linear representation capabilities of autoencoders to address the challenges of high-dimensional, non-linear data in DER networks. A comparative analysis of undercomplete and sparse autoencoders demonstrates that a sparse model utilizing L1 regularization achieves a superior F1-score of 0.95 while requiring only 20% of available training data, highlighting model efficiency and robustness to data scarcity. Furthermore, the dissertation evaluates coordinated distributed denial of service (DDoS) attacks in a real-time, multi-node edge testbed. Findings indicate that a variational autoencoder (VAE) provides robustness against stealthy, low-volume variants (precision of 0.87 - 0.91), while deterministic undercomplete and sparse autoencoder models remain more reliable for detecting high-intensity pulses.
This work concludes that a hybrid defense strategy i.e., integrating a deep learning solution that is capable of handling stealthy and overt network indicators is essential for securing DER aggregators. These findings provide an empirical basis for hardening critical energy infrastructure against sophisticated, coordinated cyber-physical attacks.
Recommended Citation
Sugunaraj, Niroop, "A Multi-Level Deep Learning-Based Approach To Hierarchical Distributed Energy Resource Management System (DERMS) Security" (2026). Theses and Dissertations. 13983.
https://commons.und.edu/theses/13983