Data-Driven Methods for Distribution Modeling and Sensing

Marković, Marija S.

Graduate Thesis Or Dissertation

Data-Driven Methods for Distribution Modeling and Sensing Public Deposited

Analytics

Download PDF

Citations

Citeable URL: https://scholar.colorado.edu/concern/graduate_thesis_or_dissertations/0k225c739

Abstract

The electric power industry has set critical renewable energy targets to promote low-emission growth in power supply. However, the integration of variable and uncertain renewable energy sources into the power grid adds complexity to its operations. This complexity is further amplified by the emergence of “grid-edge” technologies such as home battery storage, electric vehicles, and behind-the-meter renewables like rooftop solar. Collectively referred to as distributed energy resources (DERs), these clean energy technologies emphasize the importance of modernizing grid operations to ensure reliability and quality of power supply, particularly at the grid edge. But with inadequate tools to cope with limited distribution grid visibility, modernizing its operating practices is a challenge. This calls for the development of specialized tools that can operate effectively despite the lack of reliable network models and abundant measured data. This dissertation focuses on the design and study of data-driven algorithms that enable the realization of such tools, thereby contributing to data-driven methods for distribution modeling and sensing.

The dissertation is divided into two distinct parts, each targeting a specific challenge related to distribution voltage determination. The first part (Chapter 2 through Chapter 4) considers data-driven approaches for enhancing present and look-ahead situational awareness of distribution gridvoltages to effectively manage the increasing presence of DERs. Previous efforts to improve visibility into distribution system states frequently assume that the system model is fully known. In reality, however, electric distribution utilities usually lack confidence in the accuracy of these models, which limits the practicality of such solutions. In response to this, Part I of the thesis explores alternative methods that do not require knowledge of the network model. Two promising data-driven methods have been identified: machine learning (ML) and matrix completion techniques. The thesis studies both methods, focusing on their applications to model-free voltage estimation under low observability conditions (Chapter 2 and Chapter 3) and model-free voltage prediction with uncertainty estimates (Chapter 4).

The second part of the thesis (Chapter 5) shifts its focus from model-free approaches to power flow modeling for high-fidelity calculation of voltage profiles in radial distribution configurations. Motivation lies in increasing the accuracy of voltage solutions over a wider range of operating points while preserving the mathematical simplicity of the linear power flow model. To accomplish this, Part II of the thesis introduces a distribution power flow linearization that combines elements of both the underlying network model and the ML methodology. Gaussian process (GP) regression is chosen as the methodology for its capacity to generate closed-form predictions with quantified uncertainties while requiring minimal data (as also explored in Chapter 4). Within this framework, a new model is proposed and evaluated, which uses a GP-based parameterization to establish a linear relationship between squared voltages and net-load (load minus renewable generation) power injections. This new model is therefore coined as a parameterized linear power flow model.

Creator