Bose, S. (ECE) – Learning-Augmented Optimization, Control, and Inference in Modern Power Systems

The electric grid is essential to modern society, and recent developments such as renewable energy sources (RESs), battery energy storage systems (ESSs), and microgrids (MGs) have necessitated novel computational methods for planning and operations. Machine learning offers a promising lever here, both as an accelerator for and proxy to traditional optimization-based problems. In this thesis, we consider learning-based algorithms for three such problems: load restoration in islanded microgrids, accelerated optimal power flow, and short-term load forecasting.

We first address load restoration of islanded MGs containing RESs, battery ESSs, microturbines, and inverter-based devices. We formulate the problem as a multi-timestep nonconvex optimization and decompose it via model predictive control (MPC). We develop novel convex relaxations of the nonconvex constraints, including power flow, ESS charge/discharge complementarity, and inverter voltage-reactive power relations, to generate approximately feasible solutions, and then improve on them via a reinforcement learning method based on constrained policy optimization (CPO) that respects the original nonconvexity.

We then turn to accelerating convexified optimal power flow (C-OPF) via constraint screening, presenting an analysis that reduces screening for certain C-OPF families to a rank-based test. Building on this, we introduce Mixture of Gradient Experts (MoGE), an architecture that learns optimal dual variables from historical C-OPF solutions and combines them with the KKT conditions to eliminate likely non-binding constraints, with a recovery step that guarantees the reduced problem’s solution matches the original’s. We demonstrate speedups on grids with up to 10,000 buses.

Finally, we consider short-term load forecasting (STLF) from smart-meter data, motivated by the role of forecasts as inputs to the optimization problems above. To address consumer-data privacy and the heterogeneity of consumption patterns, we introduce personalization layers for federated learning (PL-FL), in which each client trains a model with a local personalized component and a shared aggregated component, and extend it to a privacy-preserving variant (PPFL) that applies differential privacy to the shared component. Separately, we present an empirical study of forecasting architectures spanning classical recurrent networks to fine-tuned time-series foundation models, holding dataset size and parameter count constant to isolate architectural contribution. All methods are evaluated on subsets of the NREL ComStock dataset.

Event Host: Shourya Bose, Ph.D. Candidate, Electrical & Computer Engineering 

Advisor: Yu Zhang

Zoom: https://ucsc.zoom.us/j/93511298189?pwd=eAyDKdMirlVqYGUsbhQCccoBM9gDV6.1

Passcode: 462014