Forecasting of Solar Eruptions using Statistical Mechanics, Ensemble, and Bayesian Forecasting Methods

Heron Systems proposes Solar Prediction via Deep Learning (SPINDLE), a human out-of-the-loop system to improve the state of solar flare forecasting using novel machine learning techniques. Currently, solar flare forecasting is either dependent on an expert, with their own subjective biases and intuitions, or automated methods using shallow representations extracted from magnetogram images, unable to learn deeper relationships in the data. SPINDLE is an automated deep learning pipeline designed to perform state-of-the-art analysis on solar observatory data for the purpose of solar flare prediction. Magnetogram and other data are collected from observatories, pre-processed, and then fed into the deep learning prediction pipeline for classification of X, M, and C solar flares in 6, 12, and 24 hour time windows. Deep learning enables the system to automatically learn sub-structures within image data over time-series, with the potential to not only dramatically improve forecasting itself, but also advance our understanding of the underlying mechanisms in solar flares. To demonstrate feasibility, we will benchmark the system against NOAA forecasts, as well as reported results in the solar flare machine learning literature.