Optimizing machine learning for space weather forecasting and event classification using modified metaheuristics

Space weather profoundly impacts Earth and its surrounding space environment, necessitating improved prediction to safeguard critical infrastructure such as communication and satellites. Solar flares can disrupt communications and pose radiation risks to airline passengers. While traditional methods offer rough estimates of solar activity trends, the potential of artificial intelligence in this domain warrants exploration. This study addresses this research gap by evaluating the performance of recurrent neural networks (RNNs) for sunspot forecasting and assessing the suitability of extreme gradient boosting (XGBoost) for solar event classification. Two publicly available datasets serve as the foundation for this research. To enhance algorithm performance through optimal hyperparameter selection, metaheuristic optimizers are employed. A unique contribution is the introduction of a modified particle swarm optimization algorithm, specifically tailored to the study's needs. Two experiments were conducted: In the first, RNNs predicted sunspot occurrence up to three steps ahead. The best-performing model, optimized using the introduced modified metaheuristic, achieved an impressive R-2 value of 0.840448, surpassing competing algorithms. In the second experiment, XGBoost models assessed solar flare severity, with the top model again optimized by the modified metaheuristic, achieving an accuracy of 0.981565. This novel approach highlights the potential for enhancing solar activity forecasting techniques and offers valuable insights into feature impacts on model decisions, thereby advancing our understanding of space weather.