Integrated Learning and Optimization and Computationally Efficient Model Predictive Control for Power Systems Applications

Overview

Power systems are undergoing a rapid transformation due to the increasing integration of renewable energy resources, distributed generation, and data-driven operational strategies. These developments require computationally efficient optimization and control techniques capable of supporting reliable real-time decision-making while accounting for uncertainty in system parameters. This thesis develops novel methodologies for model predictive control (MPC) and integrated learning and optimization (ILO) to improve the computational efficiency, robustness, and operational performance of modern power systems.

The first part of the thesis develops a computationally efficient closed-form MPC framework for primary-level voltage control of hybrid energy storage systems in renewable microgrids. The proposed formulation preserves the solution quality of conventional active-set MPC while substantially reducing computational and memory requirements. The reduced computational complexity enables the controller to satisfy short sampling intervals, improves scalability for larger storage systems, decreases power consumption, and facilitates implementation on low-cost embedded microcontrollers without sacrificing stability or robustness.

The second part of the thesis investigates integrated learning and optimization for electricity market applications involving Economic Dispatch (ED) and Direct Current Optimal Power Flow (DCOPF). The work identifies optimization parameters that are unavailable during problem solving and require prediction, including load demand and the power transfer distribution factor (PTDF) matrix, and analyzes how prediction errors propagate through optimization to influence dispatch decisions, feasibility, congestion, and market operating costs. The interactions among multiple unknown parameters are examined, and a theoretical hypothesis is developed to distinguish optimality and sub-optimality regions for parameter training, providing new insights into the relationship between prediction errors, feasibility, and decision quality.

Building on these theoretical foundations, the thesis develops a novel integrated learning and optimization framework for joint prediction of load demand and PTDF parameters. Unlike conventional sequential prediction pipelines that optimize prediction accuracy, the proposed framework incorporates the underlying optimization problem directly into neural network training through a newly derived regret formulation and its corresponding gradient. This enables the prediction model to minimize post-optimization operational regret while maintaining feasibility and capturing the operational characteristics of real-time electricity markets.

Extensive computational studies demonstrate that the proposed methodologies consistently improve computational efficiency, scalability, robustness, and decision quality. The proposed MPC framework achieves significant reductions in computational complexity while maintaining control performance, whereas the integrated learning and optimization framework reduces real-time market correction costs, improves congestion management, and produces superior operational decisions compared with conventional sequential learning approaches. Collectively, the contributions establish new computational and learning methodologies for efficient real-time control and optimization in modern power systems.