Deep learning for the prediction of geomagnetic events

In the last decades, the rising Big Data revolution together with the consolidation of High Computing capacity led to the evolution of computer systems, able now to perceive their environment thanks to the extraction of meaningful information, developing learning capacities to be self-adaptable. The depicted context has laid the perfect foundations of "Machine Learning", a subfield of Artificial Intelligence, considered at the forefront for the construction of a smart society. Its spreading is being affected all domains, finding application in medicine, transport, environment, and all industrial sectors, with the objective not only of automatizing static activities performed by human beings but also supporting them in taking decisions for more complex and dynamic challenges, capturing valuable knowledge in some cases not directly human-accessible. In such digital enhancement, neural networks are becoming more and more popular as an ML technique since, in case of a sufficient amount of information and computation capacities, they result effective whichever the field. Characterized by a simple implementation phase, they ensure flexibility in structure and configuration, adopting a suited version according to the peculiarities of the problem. Furthermore, their learning process does not require physical knowledge about the phenomenon of interest. A challenging task for these algorithms is represented by the Forecasting of specific events: the ML tool may leverage the previous history characterizing a phenomenon with the purpose of predicting its future behaviors. Mostly used in the weather and market domain, this implementation may result very useful also in the space weather field, as the prediction of specific geomagnetic events affecting the Earth may help in preventing harmful consequences. The goal of this thesis work is that of implementing neural networks for a specific class of geomagnetic events, called Coronal Mass Ejection, trying to predict future trends and classify their severity based on a time series dataset containing measurements taken in situ L1. The project has been carried out in collaboration with the Osservatorio Astronomico di Torino (OATo), here interested in obtaining in advance future information about the strength of the ring current around Earth caused by solar protons and electrons (DST). In the following work two main phases may be distinguished: at the beginning, analysis and manipulation of the dataset have been performed, trying to catch its characteristics and adopting augmentation techniques to transform it to be ready for the prediction step. Then the implementation of ML techniques has been deployed, where multiple NN architectures and settings were tested for different prediction tasks, so to prove the effectiveness of deep learning algorithms in reaching the desired goal.