Date of Award

8-17-2024

Document Type

Dissertation

Abstract

Geomagnetically induced currents (GICs) are capable of damaging existing infrastructure on the ground with the fear of prolonged and extensive global damage. We aim to predict when and where these GICs may occur with a lead time sufficient to provide preventative actions to mitigate the damage of GICs. We utilize different machine learning algorithms, as part of the Machine-learned Algorithms for GICs in Alaska and New Hampshire (MAGICIAN) team, to create both localized and global predictions of the geomagnetic field from solar wind and interplanetary magnetic field data. In my first paper, I utilize Long-Short Term Memory (LSTM) neural networks with real time magnetometer data to predict the north-south component of the geomagnetic field at 4 loca¬ tions perpendicular to the auroral oval. In my second paper, I use 7 stations of data and split each dataset based on the station’s location in magnetic local time to create 168 unique datasets across the northern hemisphere from 50o to 90o magnetic latitude, creating a global predictive network using convolutional neural networks (CNNs). In my last paper, I utilized a gridded dataset with a residual neural network (ResNet) to create predictions at 600 grid points from 40o to 90o magnetic latitude for a global model enhanced beyond the capabilities of the second paper. Machine learned models are compared to current state of the art empirical models and real data for validation. We find that LSTM networks provide good predictions at individual locations in Alaska, however, do not sup¬ port gridded global predictions with our data input scheme. In the second and third papers we find CNN and ResNet techniques predict GICs more accurately than similar empirical models without smoothing or averaging of the input features. This indicates that these machine learning techniques have the potential to provide insight on which transient solar wind features have an impact on the ground geomagnetic field perturbations.

Handle

http://hdl.handle.net/11122/15459

Download

Share

COinS