DETECTION OF SOLAR FLARES FROM IONOSPHERIC DATA USING DEEP LEARNING

This work addresses the detection of solar flares based on data from low-frequency radio wave signal amplitudes. The work is a cooperation between the Technical University of Košice and the Institute of Physics at the University of Belgrade. The primary goal is to explore the provided data and to look for patterns and connections in the ionosphere’s behavior during normal conditions and solar flares. To achieve this goal, an Autoencoder neural network is implemented in Python. Data exploration utilizes various types of data visualizations, including box plots, violin plots, and histograms. The data used included VLF signal amplitudes received in Belgrade and information on solar flares from NASA GOES satellites. Working with an extremely unbalanced dataset, low class distinctiveness, and noisy data presented significant challenges. Despite these difficulties, the final Autoencoder model, trained on non-flare data and using two key features derived from signal statistics, was able to capture 65% of real flare occurrences, although it produced many false positives. The results of this research work demonstrate that subtle differences in the signals related to flares exist and provide a basis for future efforts to improve detection.