University of Nevada - Reno
Power and Resource-Aware FPGA Acceleration of Recurrent Neural Network Controllers for Power Electronics Converters
Abstract
dc:description.abstractThe increasing deployment of inverter-based resources in modern power systems has intensified the demand for high-performance real-time control of power electronic converters. Recurrent Neural Network (RNN) controllers have demonstrated superior performance over conventional Proportional-Integral-Derivative (PID) controllers for power converter applications, yet their deployment on embedded hardware remains a significant challenge. While Digital Signal Processors (DSP) have served as the conventional platform for embedded power converter control, their limited computing power and resources present a barrier when Neural Network (NN) controllers are deployed — constraints that manifest as sequential instruction execution, instruction pipeline bottlenecks, and memory bandwidth limitations. These architectural constraints establish Field-Programmable Gate Arrays (FPGA) as they support massive parallelism, and deterministic high-frequency operation — as the natural platform for NN based power converter control. However, realizing NN controllers on FPGAs within the efficient resource, power, and latency budgets of embedded platforms remains an open and consequential design challenge.This dissertation presents a power and resource-aware FPGA acceleration framework for RNN controllers targeting real-time power electronics converter control. A progressive hardware implementation methodology is developed and validated on a two resource-constrained Intel FPGA platform, advancing from a floating-point feasibility baseline through fixed-point optimization, power-aware activation function design, and physical hardware verification, culminating in a Chebyshev polynomial based architecture that advances the resource, power, and latency boundaries of NN hardware deployment for power converter control. The contributions encompass a complete FPGA realization of a RNN controller for real-time power converter control, a power and resource-aware comparative evaluation of activation function architectures on a common industry-grade platform, resource-efficient fixed-point designs with zero on-chip memory overhead, physical board-level verification confirming practical deployability under real hardware constraints, and a Chebyshev polynomial approximation architecture achieving superior resource efficiency, power optimization, and computational latency with high numerical accuracy — confirming that power and resource-aware FPGA acceleration of RNN controllers is practically achievable for real-time power electronics converter applications.
Degree
thesis:*- Level thesis:degree_level
- Doctorate Degree
- Department thesis:degree_department
- Electrical Engineering
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Hingu, Chanakya
- Advisor dc:contributor.advisor
-
- Fu, Xingang
- Committee members dc:contributor.committeemember
-
- Fajri, Poria
- Xu, Hao
- Zhu, Xiaoshan
- Hu, Rui
Subjects
dc:subject × 4Rights
- Language dc:language.iso
- en_US, English
Identifiers
dc:identifier.*- Repository record dc:identifier.uri
- https://scholarwolf.unr.edu/handle/11714/11882