Doctoral Dissertations

Keywords and Phrases

Active Power Sharing; Deep Reinforcement Learning; Microgrids; Reactive Power Sharing; Virtual Inertia and Virtual Damping; Virtual Synchronous Generator

Abstract

The integration of inverter-based distributed generators (IBDGs) in modern power systems has ushered in a new era of renewable energy utilization, prompting the rise of Microgrid (MG) architectures. Nevertheless, as IBDG penetration increases, a host of challenges surface. Foremost among these challenges are issues related to frequency stability stemming from the inherent lack of inertia in IBDGs, and inaccurate sharing of reactive and active power, particularly evident in parallel IBDG networks with mismatched feeder impedance. This research presents novel solutions to these challenges by adopting deep reinforcement learning (DRL), specifically, the twin delayed deep deterministic policy gradient (TD3) algorithm, to achieve model-free control design. For the first challenge, an adaptive virtual synchronous generator (VSG) is developed to enhance MG stability by emulating inertia and damping. The TD3 agent ensures bounded frequency response and rate of change of frequency (ROCOF) within specified limits while preserving quick settling time. Addressing the inaccurate reactive power sharing and voltage regulation, two novel strategies are introduced: a single-agent and a multi-agent approach. The latter, employing decentralized training and execution, mitigates reactive power-sharing errors (RPSE) while ensuring voltage stability and reduced circulating current in the MG. The third challenge relates to frequency restoration and precise active power sharing when parallel VSGs operate in autonomous mode and their connecting impedances are mismatched. Therein, two novel solutions are proposed: a single-agent and a multi-agent approach. The latter allows for simultaneous regulation of system frequency and accurate active power sharing (APS) through decentralized training with shared states for each multi-agent DRL agent. Key to the success of these strategies is the crafting of reward functions that accurately reflect the desired objectives. Validation and testing are conducted using MATLAB/Simulink and Real-Time Digital Simulator (RTDS) across various MG architectures.

Advisor(s)

Kimball, Jonathan W.
Bo, Rui

Committee Member(s)

Shamsi, Pourya
Ferdowsi, Mehdi
McMillin, Bruce M.

Department(s)

Electrical and Computer Engineering

Degree Name

Ph. D. in Electrical Engineering

Publisher

Missouri University of Science and Technology

Publication Date

Summer 2024

Pagination

xiii, 152 pages

Note about bibliography

Includes_bibliographical_references_(pages 50, 86, 113 and 149-161)

Rights

©2024 Oroghene Oboreh-Snapps , All Rights Reserved

Document Type

Dissertation - Open Access

File Type

text

Language

English

Thesis Number

T 12445

Download

Share

 
COinS