An operational approach to forecast the Earth’s radiation belts dynamics

Guillerme Bernoux,Miho Janvier,Éric Buchlin,Angélica Sicard,Antoine Brunet

doi:10.1051/swsc/2021045

Abstract

The Ca index is a time-integrated geomagnetic index that correlates well with the dynamics of high-energy electron fluxes in the outer radiation belts. Therefore, Ca can be used as an indicator for the state of filling of the radiation belts for those electrons. Ca also has the advantage of being a ground-based measurement with extensive historical records. In this work, we propose a data-driven model to forecast Ca up to 24 h in advance from near-Earth solar wind parameters. Our model relies mainly on a recurrent neural network architecture called Long Short Term Memory that has shown good performances in forecasting other geomagnetic indices in previous papers. Most implementation choices in this study were arbitrated from the point of view of a space system operator, including the data selection and split, the definition of a binary classification threshold, and the evaluation methodology. We evaluate our model (against a linear baseline) using both classical and novel (in the space weather field) measures. In particular, we use the Temporal Distortion Mix (TDM) to assess the propensity of two time series to exhibit time lags. We also evaluate the ability of our model to detect storm onsets during quiet periods. It is shown that our model has high overall accuracy, with evaluation measures deteriorating in a smooth and slow trend over time. However, using the TDM and binary classification forecast evaluation metrics, we show that the forecasts lose some of their usefulness in an operational context even for time horizons shorter than 6 h. This behaviour was not observable when evaluating the model only with metrics such as the root-mean-square error or the Pearson linear correlation. Considering the physics of the problem, this result is not surprising and suggests that the use of more spatially remote data (such as solar imaging) could improve space weather forecasts.

Highlights

One of the current main topics of interest in the space weather field is forecasting geomagnetic indices based on machine learning methods
We can see that the classical metrics (RMSE, R) give much better values with the Long-Short Term Memory network (LSTM)-NN model than the linear baseline
The root-mean-square error (RMSE) values should be put into perspective with the statistical distribution of the Ca index, which over the test period has a variance of 8.9 nT and an interquartile range of 10.5 nT

Summary

Introduction

One of the current main topics of interest in the space weather field is forecasting geomagnetic indices based on machine learning methods. Machine learning has allowed for a great improvement in short-term forecasts of geomagnetic indices such as the global index Kp (Wintoft et al, 2017; Tan et al, 2018; Chakraborty & Morley, 2020) or Dst index (Gruet et al, 2018; Lethy et al, 2018). That is why the reliable forecast of geomagnetic indices and other spaceweather relevant physical quantities (e.g., relativistic electron or proton fluxes in the radiation belts) is of paramount importance. Being able to accurately and reliably forecast the fluxes of high-energy electrons (from dozens of kiloelectronvolts to a few megaelectronvolts) in the radiation belts would represent a great leap towards better mitigation of the radiation-induced risks in space. A considerable review of the methods used to forecast these electron fluxes was recently proposed by Camporeale (2019), where it is detailed that feed-forward neural networks and recurrent neural networks (RNNs) are used to obtain forecasts up to a few hours or a few days ahead (see e.g., Ling et al, 2010; Wei et al, 2018)

Results

Discussion

Conclusion