Date of Award:


Document Type:


Degree Name:

Master of Science (MS)


Computer Science

Committee Chair(s)

Soukaina Filali Boubrahimi


Soukaina Filali Boubrahimi


Nicholas Flann


Curtis Dyreson


Solar flare events are explosions of energy and radiation from the Sun’s surface. These events occur due to the tangling and twisting of magnetic fields associated with sunspots. When Coronal Mass ejections accompany solar flares, solar storms could travel towards earth at very high speeds, disrupting all earthly technologies and posing radiation hazards to astronauts. For this reason, the prediction of solar flares has become a crucial aspect of forecasting space weather. Our thesis utilized the time-series data consisting of active solar region magnetic field parameters acquired from SDO that span more than eight years. The classification models take AR data from an observation period of 12 hours as input to predict the occurrence of flare in next 24 hours. We performed preprocessing and feature selection to find optimal feature space consisting of 28 active region parameters that made our multivariate time series dataset (MVTS). For the first time, we modeled the flare prediction task as a 4-class problem and explored a comprehensive set of machine learning models to identify the most suitable model. This research achieved a state-of-the-art true skill statistic (TSS) of 0.92 with a 99.9% recall of X-/M- class flares on our time series forest model. This was accomplished with the augmented dataset in which the minority class is over-sampled using synthetic samples generated by SMOTE and the majority classes are randomly under-sampled. This work has established a robust dataset and baseline models for future studies in this task, including experiments on remedies to tackle the class imbalance problem such as weighted cost functions and data augmentation. Also the time series classifiers implemented will enable shapelets mining that can provide interpreting ability to domain experts.