Forecasting Solar Flares Using Shallow and Deep Learning Rechniques
Solar flares, which are large sudden increases in X-ray flux, can damage satellite infrastructure, hamper power grids, disrupt Global Positioning Systems (GPS), and impair long-distance communication. Thus, the accurate prediction of solar flares has high practical importance and numerous approaches to the problem have been proposed. While these works have shown promising results, they have also highlighted the inherent difficulty of the problem, as solar flares are rare and heterogeneous events and the physical phenomena that give rise to them are still not well-understood. Solar flare prediction is normally posed as a classification problem, where a sequence of measurements is classified as a precursor or not of a solar flare within a given time frame. In this dissertation, we implemented and evaluated multiple algorithms for solar flare prediction posing the problem as a regression problem, focusing on the prediction of the maximum flux within a fixed future time interval. We compared conventional machine learning algorithms, state-of-the-art deep learning architectures tailored to time series regression, and a recently-proposed randomized shallow model called ROCKET, that has shown excellent results in other applications. Our experimental results show that ROCKET outperforms all other algorithms in terms of mean-squared error, while requiring much shorter training times than deep neural networks. Taking advantage of its efficiency, we propose an ensemble of ROCKET models, which leads to a further improvement in results.
Astrophysics|Artificial intelligence|Computational physics
Dey, Sumi, "Forecasting Solar Flares Using Shallow and Deep Learning Rechniques" (2022). ETD Collection for University of Texas, El Paso. AAI29326601.