Interpretable Solar Flare Forecasting via Matrix Profile Shapelet Mining

Abstract Severe solar flare events can result in radio and electrical grid blackouts and damage to satellites. To protect critical infrastructure on Earth and in orbit, solar flare events must be predicted reliably. Deep learning models have been effective for solar flare prediction, but the black-box nature of deep learning makes extracting meaningful physics from such models difficult. We propose a solar flare prediction approach based on time-series shapelets, exploiting the matrix profile for efficient shapelet mining. Our aim is to demonstrate that flare prediction models based on time-series shapelets can make accurate predictions and provide meaningful physical insights into solar flare genesis. We mine time-series shapelets from each of the 24 magnetic field parameters given in the SWAN-SF dataset. We then evaluate the quality of the mined shapelets, and leverage the shapelets to create an interpretable solar flare prediction model. We demonstrate that this model is capable of predicting flare events with a true skill statistic comparable to state-of-the-art models.