Speaker
Description
Solar flares, sudden bursts of electromagnetic energy originating from magnetically active regions on the solar surface, pose significant risks to satellite infrastructure, communication systems, and power grids. Predicting these events is crucial for both physicists and governmental agencies. Recent observations and research have revealed that solar flares are more complex phenomena than previously thought, as they extend from the corona to the lower photosphere, underlining the interconnected nature of the Sun’s atmospheric layers and emphasizing the need for a comprehensive understanding of solar flare dynamics to improve space weather forecasting and our ability to protect space-based technology and infrastructure. Conventional approaches to this problem rely on features extracted from line-of-sight (LoS) magnetograms of solar active regions, which have traditionally been linked to increased flare activity. More recent methodologies leverage temporal series of LoS magnetograms to extract as much information as possible from available data, progressing towards a fully automated flare forecasting system. Despite these advancements, recent studies with LoS magnetograms haven't shown significant improvements over previous methods, raising doubts about their effectiveness. We propose a deep learning-based approach to address this challenge. We model the solar flare forecasting problem as a binary time series classification task using line-of-sight (LoS) magnetograms obtained from the Solar Dynamics Observatory's Helioseismic and Magnetic Imager (SDO/HMI). These magnetograms provide crucial data capturing magnetic field variations on the solar surface, which are essential for predicting solar activity. Our primary objective is to distinguish between active regions likely to produce M- or X-class flares within a 24-hour window and those remaining inactive. Such a classifier could be instrumental in the development of an early warning system for solar flares. Our proposed approach consists of two phases: firstly, a Convolutional Neural Network (CNN) autoencoder performs feature extraction from the magnetograms. Secondly, a Long Short-Term Memory (LSTM) binary classifier processes the extracted features to predict flare activity. Our proposed methodology achieves a remarkable 90% test accuracy. It is validated against test case AR HARP 377, demonstrating its efficacy in solar flare prediction.
