Exploring Predictive Capabilities of GOES and SDO/EVE Data for Flare Forecasting

We explore the predictive capabilities of full disk integrated solar X-ray and extreme ultraviolet emission data from the sun, collected by the Geostationary Operational Environmental Satellites and the EVE/ESP instrument on the Solar Dynamics Observatory (SDO) in view of forecasting solar flares. A purely data driven approach is studied which uses suitably preprocessed time sequence data of fixed length. Models of different complexity are tested - ranging from a model that just uses the per-sequence average up to models that use several layers of Long-Short Term Memory neural networks which are designed to capture details about the sequential nature of the data. A True Skill Statistics of roughly 60% is obtained even by a very simple model based on the per-sequence average of GOES data. This performance can not be significantly outperformed by more complex models or by incorporating EVE/ESP data from SDO. Further work would be needed to gain more evidence whether the considered full disk data contains sufficient predictive capacity for improving the performance for flare forecasting.