Sequence Model-based End-to-End Solar Flare Classification from Multivariate Time Series Data

Abstract

Machine learning-based prediction of solar flares has become an important application of data science in space weather research. Spatiotemporal magnetic field data of solar active regions captured by solar imaging observatories are mapped into multivariate time series data to facilitate temporal window-based solar flare prediction. Existing methods of solar flare prediction leveraging multivariate time series data rely heavily on statistical features for the representation learning of individual time series instances. In this work, we used Deep Learning, more specifically Long Short Term Memory (LSTM) Networks for learning representations of multivariate time series instances that map into multiple flare classes. This work enables the end-to-end multivariate time series classification bypassing the requirements of hand-engineered features. Our experiments on a real-life solar flare dataset show better prediction performance in comparison with state-of-the-art baseline methods.