Solar Spectrum – Understanding Solar Activity Patterns Within a Solar Cycle Using Machine Learning Techniques

The Sun's overpowering impact on our planet and space behavior is undeniable. Solar cycles are characterized by variations in the number and distribution of sunspots, and other solar activities like solar flares and coronal mass ejections (CME). Solar cycle prediction involves forecasting the overall behavior and activity levels of the Sun over the course of several solar cycles. These activities play a crucial role in shaping space weather, influencing satellite communications, power systems, and Earth's climate. Consequently, accurately predicting and classifying solar phenomena using machine learning methods has garnered significant attention in astrophysics. This research compares the efficiency of Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) architectures in forecasting solar activities, focusing on sunspots, solar flares, and CMEs. The study aims to identify the superior model for the Solar Activity Prediction System, by employing advanced machine learning techniques. Sunspot prediction, based on a univariate dataset, revealed minimal overfitting for both models, with LSTM showing closer alignment with validated data and GRU producing a consistent average. Multivariate solar flare prediction showcased varied accuracies, with LSTM at 72.76% and GRU at 71.43%, while combined models reached up to 83.89%. Additionally, the binary classification of CMEs using LSTM and GRU models resulted in an AUC of 55 percent and 47 percent, respectively. These findings offer significant perspectives into the efficacy of Recurrent Neural Network architectures in solar activity forecasting, useful for advancing solar activity forecasting systems.