Date of Award

Spring 2021

Document Type

Open Access Dissertation


Electrical Engineering

First Advisor

Andrea Benigni


In recent years, switching power electronic systems for use in power conversion, generation, and distribution have become increasingly larger and complex in their design. With the increase in complexity comes the increase in difficulty in performing analysis, experiments, and health management on such systems in physical space. Part of this difficulty comes from challenge of access to and understanding of the internal interplay of the numerous components, safety and cost of system operation, and the inherent uncertainty of physical characteristics of systems and their environment.

A recent approach for testing and diagnostics/prognostics of power electronic systems is the application of the digital twinning concept already used in other fields. This approach involves creating a digital twin replica of a physical system which reflects the same characteristics and behavior of the physical twin counterpart. These digital twins are often simulated models of the physical counterpart, developed from data-driven characterization of the physical system to ensure it is consistently reflected by the digital twin. Digital twins can be used in place of the physical counterpart in analysis and testing, along with being used as a real-time reference model for predicted behavior of the physical twin for diagnostics or prognostics. From these features, use of digital twins can overcome many of the difficulties in working directly with physical systems.

Physical power electronic systems tend to operate under a degree of uncertainty, due to the limited knowledge of system structure and random environmental aspects on said systems. Without considering the uncertainty of a physical system within a digital twin, the digital twin cannot accurately reflect system behavior within probabilistic intervals and confidence. In this regard, the digital twins should then be developed as probabilistic models which considers both the deterministic behavior and uncertainty of the physical twins. To allow the probabilistic digital twins to serve as emulation of power electronic systems for online testing and health management, these digital twins should be real-time solvable. However, probabilistic models are computationally expensive, leading to difficulty in developing real-time solvers of such models while achieving reasonable model fidelity and scalability.

To address the need for real-time simulation of probabilistic digital twins of power electronic systems, this dissertation presents novel implementations of real-time probabilistic model solvers. These solvers utilize generalized Polynomial Chaos Expansions (PCE) to express uncertain/random processes of models in analytical form, along with leveraging Field Programmable Gate Array (FPGA) execution, to achieve computational speedups for real-time execution. The presented real-time solvers are developed as either: custom model-specific designs implemented in C++ utilizing a specially created PCE library and then high level synthesized into a FPGA hardware core; or as programs executable on a FPGA-based, PCE-tailored vector processor that was developed for this work. For both solver implementations, performance and resource usage analysis was performed to determine how these solvers accelerate real-time solving of probabilistic PCE digital twin models of switching power converters. This analysis shows that such models with handful of uncertainty sources can be simulated in real-time by the proposed solvers with time steps in nanosecond and microsecond range. An application of the presented solvers for controller-embedded, comparative diagnostics of a power converter is shown within the dissertation. Also, a rudimentary example is presented of a data-driven digital twin of a physical power converter that can be executed by the proposed real-time solvers.