Predicting Geostationary 40–150 keV Electron Flux Using ARMAX (an Autoregressive Moving Average Transfer Function), RNN (a Recurrent Neural Network), and Logistic Regression: A Comparison of Models

• Published in: Space Weather Vol. 21; no. 5
• Main Authors: Simms, L. E., Ganushkina, N. Yu, Kamp, M., Balikhin, M., Liemohn, M. W.
• Format: Journal Article
• Language: English
• Published: Washington John Wiley & Sons, Inc 01.05.2023; Wiley
• Subjects: Algorithms; ARMAX; Autoregressive moving-average models; Electron flux; electron flux prediction; Electrons; Fluctuations; Forecasting; Geomagnetic field; Geosynchronous orbits; Interplanetary magnetic field; logistic regression; Magnetic fields; Model accuracy; Modelling; Neural networks; Parameters; precision recall curve; Prediction models; Radiation; Radiation belts; Radiation damage; recurrent neural network; Recurrent neural networks; Regression analysis; Regression models; ROC curve; Solar magnetic field; Solar wind; Statistical analysis; Transfer functions
• ISSN: 1542-7390; 1539-4964; 1542-7390
• DOI: 10.1029/2022SW003263

We screen several algorithms for their ability to produce good predictive models of hourly 40–150 keV electron flux at geostationary orbit (data from GOES‐13) using solar wind, Interplanetary Magnetic Field, and geomagnetic index parameters that would be available for real time forecasting. Value‐predicting models developed using ARMAX (autoregressive moving average transfer function), RNN (recurrent neural network), or stepwise‐reduced regression produced roughly similar results. Including magnetic local time as a categorical variable to describe both the differing levels of flux and the differing influence of parameters improved the models (r as high as 0.814; Heidke Skill Score (HSS) as high as 0.663), however value‐predicting models did a poor job at predicting highs and lows. Diagnostic tests are introduced (cubic fit to observation‐prediction relationship and Lag1 correlation) that better assess predictions of extremes than single metrics such as root mean square error, mean absolute error, or median symmetric accuracy. Classifier models (RNN and logistic regression) were equally able to predict flux rise above the 75th percentile (HSS as high as 0.667). Logistic regression models were improved by the addition of multiplicative interaction and quadratic terms. Only predictors from 1 or 3 hr before were necessary and a detailed description of flux time series behavior was not needed. Stepwise selection of these variables trimmed non‐contributing parameters for a more parsimonious and portable logistic regression model that predicted as well as neural network‐derived models. We provide a logistic regression model (LL3: LogisticLag3) based on inputs measured 3 hr previous, along with optimal probability thresholds, for future predictions. Plain Language Summary As high levels of electrons in the radiation belts can damage satellites, accurate forecasting would be a useful tool. Electron levels can be predicted using information from the solar wind, the interplanetary magnetic field, and indices measuring disturbances in Earth's magnetic field. We compare several algorithms to produce such models: regression and neural networks that depend on predictors at one or many previous time steps. We find that dependable predictions can be made from a regression model using predictors from only a single previous time step. More sophisticated neural network techniques are not necessary if interaction and nonlinear terms are introduced to the regression. Key Points Regression models incorporating interaction and quadratic terms predict electron flux as well as neural network models The description of time series behavior by autoregressive moving average transfer function models, while useful for hypothesis testing, is not necessary for prediction Magnetic local time as a predictor improves the models by describing changing flux levels and the differing influence of parameters over the diurnal period