Document Type

Thesis - Open Access

Award Date


Degree Name

Master of Science (MS)

Department / School

Electrical Engineering and Computer Science

First Advisor

Timothy M. Hansen


convolutional neural network, federated learning, frequency measurement, inertia estimation, low inertia grids, power system dynamics


Recent updates to the IEEE 1547-2018 standard allow active participation of distributed energy resources (DERs) in power grid services with the goal of increased grid reliability and resiliency. With the rapid growth of DERs towards a low inertia converter-dominated grid, the DERs can provide fast frequency response (FFR) services that can quickly counteract the change in system frequency through inertial support. However, in low voltage grids, frequency and voltage face dynamics coupling due to a high resistance to reactance ratio and cannot be controlled separately as in the bulk electric grid. Due to the coupling effect, the control of one parameter also affects the dynamics of the other parameter. A part of this work highlights the role of DERs to provide grid ancillary services underscoring the challenges of combined voltage and frequency control in low voltage grids. Increasing penetration of renewable energy sources (RES) also decreases the power system inertia, there by affecting the stability of bulk grid. The stochastic nature of RES makes the power system inertia a time-varying quantity. Furthermore, converter-dominated grids have different dynamics compared to conventional grids and therefore estimates of the inertia constant using existing dynamic power system models are unsuitable. This work proposes a novel inertia estimation technique based on convolutional neural networks (CNN) that use local frequency measurements. The model uses a non-intrusive excitation signal to perturb the system and measure frequency using a phase-locked loop. The estimated inertia constants, have significant accuracy for the training, validation, and testing sets. Additionally, the proposed approach can be applied over traditional inertia estimation methods that do not incorporate the dynamic impact of renewable energy sources. The frequency response of power systems changes drastically when multi-area power systems with interconnected tie-lines are considered. Furthermore, higher penetration of RES increases the stochasticity in interconnected power systems. Hence, it is important to estimate the multi-area parameters ensuring communication and coordination between each of the areas. A robust and secure client-server-based distributed machine learning framework is used to estimate power system inertia in a two-area system. The proposed approach can be efficiently optimized to increase the training performance. It is important to analyze the performance of a trained machine learning model in a real-world scenario with unknown dynamics. A pre-trained CNN is tested on a system with model predictive controller (MPC)-based virtual inertia (VI) unit. Results show that the frequency and inertial response of conventional synchronous generators-based system differs drastically as compared to the system with non-synchronous generator-based VI support.

Library of Congress Subject Headings

Distributed resources (Electric utilities)
Machine learning.
Inertia (Mechanics)
Electric power systems -- Management.



Number of Pages



South Dakota State University



Rights Statement

In Copyright