Deep neural networks for prediction of solar flares

Abstract

Solar flares are enormous explosions on the solar surface that originates from sunspots, which could cause damage to satellites, power grids and radio communication systems. Having early-warning systems that could accurately predict the eruption of harmful flares are becoming more crucial, as our society grows more dependant on technology. Deep learning has brought about a new era of flare prediction tools and models, with astounding success compared to traditional techniques, although at a deficit of interpretability. In this study, we present findings and study the effects of several deep neural network architectures, optimised with different optimisation techniques and scaling strategies. We aim to create simplified models with enough capacity to accurately predict flares, using image-derived sunspot features. We also investigate the predictive ability of these features using deep learning attribution methods. We find that the task can be modelled such that the basic multilayer perceptron (MLP), with a small architecture, could accurately predict solar flares on par with its larger MLP counterparts. We find that including the temporal information of past observations does not increase performance, but smoothed the probabilistic forecasts in the case of our proposed one-dimensional causal convolutional (1D-CNN) network, compared to the MLP network. We also find that the attribution methods tested were able to extract the most relevant features relating to the eruption of flares, which is consistent with other findings in the literature. The study contributes towards open questions regarding imbalanced classification and probabilistic forecasting using deep neural networks, specifically giving an informed perspective on binary classification architectures, optimisation techniques, scaling strategies and attribution methods. In the process, we provide some additional insight into a well-known dataset, and develop a deployable solar flare prediction model.