Author ORCID Identifier

https://orcid.org/0009-0001-2155-1421

Semester

Summer

Date of Graduation

2026

Document Type

Thesis

Degree Type

MS

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Sarika Khushalani Solanki

Committee Co-Chair

Anurag K. Srivastava

Committee Member

Jignesh Solanki

Abstract

Modern power systems face failures at very different timescales. Cyber-physical disturbances may emerge within seconds, while battery degradation develops over hundreds of operating cycles. Both monitoring problems share the same underlying difficulty: power systems produce measurement data in abundance, but reliably labeled examples of abnormal or degraded operation are scarce. Rare grid events are difficult to label, and battery degradation data are heterogeneous across cells, cycling protocols, and chemistries. This thesis addresses these challenges through two complementary domain-informed learning frameworks that constrain representation learning using information specific to each physical problem.

First, at the system level, T-BiGAN, a Transformer-augmented bidirectional generative adversarial network, learns the nominal spatiotemporal manifold of synchrophasor measurements across distributed phasor measurement units (PMUs) and detects departures from normal operation using a composite score built from reconstruction error, latent-space inconsistency, and discriminator confidence. Trained without anomaly labels and evaluated on a realistic hardware-in-the-loop (HIL) benchmark of an IEEE 39-bus system with eight PMUs, T-BiGAN achieves a precision of 0.90, recall of 0.81, F1-score of 0.88, ROC-AUC of 0.951, and average precision of 0.996, outperforming leading supervised and unsupervised baselines.

Second, at the asset level, QPINN, a physics-informed neural network with quantum-inspired feature mapping, estimates battery state of health (SOH) by combining a fixed Nyström-approximated quantum kernel embedding with degradation dynamics and monotonicity constraints. Validated on 310,705 samples from 387 cells spanning four datasets and multiple lithium-ion chemistries, QPINN achieves an average SOH estimation accuracy of 99.46\%, reducing MAPE and RMSE by up to 65\% and 62\% relative to state-of-the-art baselines, and supports cross-dataset adaptation using limited labeled target-domain data.

Finally, the thesis develops an equivalent-circuit-model (ECM) parameterization and hardware-in-the-loop (HIL) evaluation workflow for non-lithium battery chemistries, intended as the modeling foundation for future utility-scale battery digital twins. A data-identifiability analysis of aluminum-air battery data shows that constant-current discharge records cannot excite the RC dynamics required for reliable ECM identification, motivating a specification of the pulse-based characterization data needed, along with an aluminum-inventory-based state-of-charge formulation. Because the present workflow relies on one-time laboratory parameterization rather than continuously streaming field data, it constitutes the model backbone of a prospective digital twin rather than an operational twin, and this limitation is stated explicitly.

Together, these contributions provide system-level and asset-level monitoring capabilities for resilient smart grids. T-BiGAN flags when the measured synchrophasor streams depart from learned normal behavior, whether the underlying cause is a physical disturbance, a data-quality problem, or a cyber event, and QPINN estimates how far a battery cell has degraded from its rated capacity.

Download

Share

COinS