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

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 real time forecasting. Value-predicting developed ARMAX (autoregressive moving average transfer function), RNN (recurrent neural network), or stepwise-reduced regression produced roughly similar results. Including magnetic local as a categorical variable describe both the differing levels influence improved (r high 0.814; Heidke Skill Score (HSS) 0.663), however value-predicting did poor job predicting highs lows. Diagnostic tests are introduced (cubic fit observation-prediction relationship Lag1 correlation) better assess predictions extremes than single metrics such root mean square error, absolute median symmetric accuracy. Classifier (RNN logistic regression) were equally able predict rise above 75th percentile (HSS 0.667). Logistic by addition multiplicative interaction quadratic terms. Only predictors 1 3 hr before necessary detailed description series behavior was not needed. Stepwise selection these variables trimmed non-contributing more parsimonious portable model predicted well network-derived models. provide (LL3: LogisticLag3) based on inputs measured previous, along with optimal probability thresholds, future predictions.