Date of Award

Spring 2026

Language

English

Embargo Period

5-1-2026

Document Type

Master's Thesis

Degree Name

Master of Science (MS)

College/School/Department

Department of Electrical and Computer Engineering

Program

Electrical and Computer Engineering

First Advisor

Nathan Dahlin

Committee Members

Mohammed Agamy , Bariscan Yonel

Keywords

Stochastic Model Predictive Control, Reinforcement Learning, Battery Energy Storage Systems, Net Load Forecasting, Scenario-Based Optimization, Renewable Energy Integration, Adaptive Energy Management, Uncertainty Quantification

Subject Categories

Controls and Control Theory | Electrical and Computer Engineering

Abstract

Battery energy storage systems play a critical role in enabling reliable operation of power systems with high penetration of renewable energy. However, optimal control of BESS is challenging due to uncertainty in net load and electricity prices. Deterministic MPC, which relies on point forecasts, can perform suboptimally under forecast errors. SMPC addresses this limitation by incorporating uncertainty through scenario-based optimization. Nevertheless, its performance depends on fixed objective parameters, particularly the cycling penalty, which governs the trade-off between economic cost and battery utilization. This thesis develops a learning-augmented SMPC framework for battery control under net load uncertainty. The proposed approach integrates machine learning-based forecasting, residual-based scenario generation using SARIMA models, and a RL agent that adaptively selects the cycling penalty in real time. This hierarchical structure enables the controller to dynamically adjust its operating regime while preserving the feasibility and constraint satisfaction properties of SMPC. The framework is evaluated using real-world data from the NYISO, including behindthe-meter solar generation, load, and day-ahead electricity prices. Results demonstrate that the SMPC formulation induces a structured trade-off between economic cost and battery throughput, and that the RL agent consistently selects operating points along this tradeoff. The learned policy achieves cost performance close to fixed-parameter SMPC while substantially reducing battery usage, thereby providing an adaptive mechanism for balancing economic efficiency and operational intensity without manual parameter tuning. Overall, the proposed approach highlights the potential of combining learning-based adaptation with model-based stochastic control. The results show that RL can effectively navigate control trade-offs and provide a practical and interpretable mechanism.

License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Download

Share

COinS