Date of Award

Spring 2025

Project Type

Dissertation

Program or Major

Electrical and Computer Engineering

Degree Name

Doctor of Philosophy

First Advisor

Md Shaad Mahmud

Second Advisor

Se Young Yoon

Third Advisor

Diliang Chen

Abstract

Geomagnetically Induced Currents (GICs) are currents induced in the Earth's surface conductors and are caused by geomagnetic disturbances or storms (GMDs). Such disturbances in the Earth's magnetic field are the result of interactions between the Earth's magnetosphere (the region of space dominated by the Earth's magnetic field) and the solar wind (a continuous stream of charged particles from the Sun) \cite{camporeale_machine_2018} \cite{geosciences12010027}. These induced currents considerably threaten vital technological systems, notably the electrical power grids. This research addresses the critical challenges of robust, real-time forecasting and monitoring of GICs using machine learning (ML). The project initiative responds to the inherent uncertainties in space weather prediction stemming from stochastic environmental behaviors and limitations in observable data. Furthermore, data from terrestrial magnetometers, which is utilized to calculate the ground horizontal magnetic component ($dB_{H}/dt$) as a proxy measure of GIC, suffers from inherent background noise (baseline). Traditional baseline correction techniques are robust but computationally demanding and unsuitable for meeting real-time requirements. Therefore, a two-phased approach is adopted to enhance the reliability and real-time capabilities of existing GIC forecasting and monitoring techniques. The first phase introduces an offline Bayesian ensemble deep learning (DL) framework for predicting $dB_{H}/dt$. This approach employs Bayesian statistics to manage model and data uncertainty, producing predictions with a 95\% credible interval. The second phase explores a branch of edge machine learning (ML) called Tiny Machine Learning (TinyML), which focuses on developing optimized models for resource-limited edge devices, like microcontroller units (MCUs). A proof-of-concept for a cost-effective magnetometer system that utilizes TinyML for real-time baseline correction and $dB_{H}/dt$ forecasting has been presented. Due to the inherent resource limitations of MCUs, they are unable to support robust yet computationally intensive models that leverage Bayesian machine learning techniques. Consequently, this research further proposes two separate strategies. The first involves a statistical model that approximates a benchmark variational autoencoder (VAE), a variant of a Bayesian DL architecture designed for baseline correction. This model also estimates the prediction variance to quantify the uncertainty. The second strategy introduces a novel optimization scheme that embeds a physics-based regularization term into the model loss function. The second strategy presents an innovative optimization framework for TinyML, incorporating physics-based regularization into the model's loss function to optimize deep learning model elements while maintaining robustness. This approach ensures that the model's adjustments align with the governing physics of solar wind and magnetosphere interactions by employing a sensitivity-driven pruning and quantization scheme. The overarching objective of the proposed methodologies is to lay the groundwork for conceptualizing advanced, ML-driven magnetometer systems capable of real-time baseline correction and forecasting geomagnetic anomalies.

Download

Share

COinS