Solar Flare Prediction With Recurrent Neural Networks

Jill Platts, AFRL/RISA; John Marsh, SUNY Polytechnic Institute; Michael Reale, SUNY Polytechnic Institute; Christopher Urban, SUNY Polytechnic Institute

Keywords: space weather, solar flares, machine learning, neural networks, recurrent neural networks

Abstract:

With our reliance on technological systems rising, in addition to the increased demand for space travel and exploration, the need to improve our ability to accurately forecast space weather continues to gain attention. The largest type of explosive event that directly impacts Earth, a solar flare is an enormous burst of radiation visible via optical light, as well as X-ray and Ultraviolet (UV). Occurring quickly, solar flares can damage satellites, disrupt power grids and expose astronauts to intense radiation and unstable conditions. Exhibiting the Sun’s overall unpredictability, solar flare events can be both abrupt and short or gradual and long. The diversity of solar flare events continues to present a forecasting conundrum.

The goal of this research is to further explore and develop strategies for solar flare prediction, specifically involving machine learning. A popular choice when working with time series data, or sequences, Recurrent Neural Networks (RNNs) are an excellent type of neural network for solar flare forecasting models. Equipped with an internal memory, RNNs are able to handle time sequences more effectively than other types of neural networks. This research aims to prove the validity of using RNNs with multivariate time series data, related to the Sun’s magnetic field, to predict solar flares.

Monitored nearly 24 hours a day by various satellites, an amalgamation of the Sun’s magnetic field properties, observed over a period of time, is theorized to be indicative of an upcoming solar flare. Springboarding off past research efforts at solar flare prediction, the open source dataset from Solar Dynamics Observatory’s (SDO’s) Helioseismic and Magnetic Imager (HMI) is used to procure and assemble a set of parameters related to the Sun’s photosphere, specifically the behavior of Active Regions (ARs). ARs are clearly visible concentrated clusters of sunspots on the Sun’s photosphere that are currently experiencing intense vertical magnetic flux. Each time sequence used for this research consists of a certain set of parameters which track AR activity over a 24-hour period. Specifically, parameters are derived from SDO’s open source Spaceweather HMI Active Region Patch (SHARP) product. The SHARP product contains various space weather quantities calculated from photospheric vector magnetogram data.

National Oceanic and Atmospheric Administration’s (NOAA’s) Geostationary Operational Environmental Satellites (GOES) yearly X-ray reports, as well as the Space Weather Prediction Center’s (SWPC’s) daily solar event reports, are used identify solar flare events by both class (based on X-ray flux) and date of occurrence. Solar flare events are linked directly to SHARP-based time sequences via an official number assigned by NOAA. The resulting dataset provides a basis for training and tuning multiple RNN model architectures. The combination of solar flare occurrences and the corresponding AR region’s behavior before, during and after time of occurrence offer an effective supervised learning approach to solar flare prediction.

Various RNN-based models are tuned to both predict and classify solar flares 24 hours in advance, with promising results. Particular focus is given to Gated Recurrent Units (GRUs). The newest form of a RNN, GRU’s have shown superior performance against both basic RNNs and Long Short-Term Memory (LSTM) RNNs. Accepted methods for analyzing prediction and classification machine learning algorithms, including confusion matrix, precision, recall, F1 score and accuracy, are used to measure the performance of model architectures against each other and against past research. Current results and expected future work indicates GRUs consistently show a superior level of prediction over basic RNNs and LSTMs.

In an effort to protect space assets, as well as operations on Earth, machine learning solutions appear to be the best tool for solar flare forecasting. Working to mitigate the harmful impact of space weather, achieving an accurate prediction window of one day provides ample time for satellite relocation and other necessary actions. Further efforts with solar flare prediction expect to improve forecasting accuracy with additional model architectures and tuning.

Date of Conference: September 14-17, 2021

Track: Atmospherics/Space Weather

View Paper