Short-Term Forecasting of the SEP Flux during the Solar Cycle 25 Using the LSTM Neural Network 

Mohamed Nedal - Institute of Astronomy of the Bulgarian Academy of Sciences; Kamen Kozarev - Institute of Astronomy of the Bulgarian Academy of Sciences, Bulgaria 

Session: QV - Interplanetary CMEs and Solar Particle Events

Abstract

We developed and trained a forecasting model using a Long Short-Term Memory (LSTM) neural network model, with the goal to make short-term predictions for SEP fluxes at three integral energy bands (>10 MeV, >30 MeV, and >60 MeV). 

Here, we trained a model with a combination of solar and interplanetary magnetic field indices obtained from OMNI database, which are the magnitude of the average interplanetary magnetic field, plasma flow speed, the F10.7 index, and the X-ray flux in two wavelengths (0.1 – 0.8 nm and 0.05 – 0.4 nm) obtained from GOES satellite database for the past four solar cycles. We generate and evaluate predictions of the hourly-averaged SEP fluxes in three integral energy bands at 1 AU for 1-, 3-, 7-day periods.

So far, we found that the Mean Absolute Percentage Errors (MAPEs) of the predictions for the 1-, 3-, and 7-day periods at the energy band >10 MeV were < 23.6%, 12.1%, and 19.7%, respectively. At the energy band >30 MeV, the MAPEs were < 23.6%, 12.2% and 12.3%, respectively. At the energy band >60 MeV, the MAPEs were < 23.9%, 18%, and 19.6%, respectively. We also found that the model underestimates the prediction at the highest integral energy band.