System Identification of Local Time Electron Fluencies at Geostationary Orbit

Abstract

The electron fluxes at geostationary orbit measured by Geostationary Operational Environmental Satellite (GOES) 13, 14, and 15 spacecraft are modeled using system identification techniques. System identification, similar to machine learning, uses input-output data to train a model, which can then be used to provide forecasts. This study employs the nonlinear autoregressive moving average exogenous technique to deduce the electron flux models. The electron fluxes at geostationary orbit are known to vary in space and time, making it a spatiotemporal system, which complicates the modeling using system identification/machine learning approach. Therefore, the electron flux data are binned into 24 magnetic local time (MLT), and a separate model is developed for each of the 24 MLT bins. MLT models are developed for six of the GOES 13, 14, and 15 electron flux energy channels (75 keV, 150 keV, 275 keV, 475 keV, >800 keV, and >2 MeV). The models are assessed on separate test data by prediction efficiency (PE) and correlation coefficient (CC) and found these to vary by MLT and electron energy. The lowest energy of 75 keV at the midnight sector had a PE of 36.0 and CC of 59.3, which increased on the dayside to a PE of 66.9 and CC of 81.6. These metrics increased to the >2 MeV model, which had a low PE and CC of 63.0 and 81.8 on the nightside to a high of 80.3 and 90.8 on the dayside.

Removing Diurnal Signals and Longer Term Trends From Electron Flux and ULF Correlations: A Comparison of Spectral Subtraction, Simple Differencing, and ARIMAX Models

“…In general, ARIMAX models do a better job of forecasting relativistic electron flux than simpler regression models (Boynton et al, 2019(Boynton et al, , 2020Simms & Engebretson, 2020). However, a transfer function model may not only result in a better fit, it can also show the cumulative effects of predictor variables.…”

“…Other approaches have involved modeling each 24 hr MLT bin separately (Boynton et al., 2020), creating a noon proxy, using various magnetospheric inputs such as Kp, Dst, AE, and ULF wave power to probabilistically generate what a flux measurement would be if it occurred at noon (O’Brien & McPherron, 2003; Su et al., 2014), asynchronous regression (O'Brien et al., 2001), or applying the Kalman filter algorithm (Rigler et al., 2004). (Despite its name, the Kalman filter is not a filter in the signal processing sense.…”

Radiation Belt Model Including Semi‐Annual Variation and Solar Driving (Sentinel)

“…to estimate the acceleration due to wave-particle interactions and thus predict the changes in electron phase space density, and so is not clear how or if they take SAV-and other large scale variations-into account. Finally, machine learning models such as the recent MERLIN model (Smirnov et al, 2020) or the Nonlinear AutoRegressive Moving Average with eXogenous inputs models (Boynton et al, 2019(Boynton et al, , 2020, built on many years of data, probably include the effects of all such variabilities, but in a way that is difficult to disentangle from all the other effects and variations.…”