Improving Solar Flare Prediction by Time Series Outlier Detection

Solar flares not only pose risks to outer space technologies and astronauts’ well being, but also cause disruptions on earth our high-tech, interconnected infrastructure lives highly depend on. While a number of machine-learning methods have been proposed improve flare prediction, none them, the best knowledge, investigated impact outliers reliability robustness those models’ performance. In this study, we investigate in multivariate time series benchmark dataset, namely SWAN-SF, prediction models, test hypothesis. That is, there exist removal which enhances performance models unseen datasets. We employ Isolation Forest detect among weaker instances. Several experiments are carried out using large range contamination rates determine percentage present outliers. assess quality each dataset terms its actual TimeSeriesSVC. findings, achieve 279% increase True Skill Statistic 68% Heidke Score. The results show that overall significant improvement can be achieved for if detected removed properly.