Comparative Analysis of TPA-LSTM and Transformer Models for Forecasting GEO Radiation Belt Electron Fluxes

Abstract

The geosynchronous orbit (GEO) is a region filled with energetic electrons and it hosts hundreds of satellites. Electron fluxes at GEO can change sharply within hours, making high-time-resolution prediction crucial. In this study, we develop and compare two neural networks for persistent high-time-resolution prediction: long short-term memory with temporal pattern attention (TPA-LSTM) and Transformer. Unlike most previous models, which only output electron fluxes, our models output the same parameters as the inputs, including magnetic local time, solar wind speed, solar wind dynamic pressure, AE, Kp, Dst, the north-south component of the interplanetary magnetic field, and electron flux data from GOES-15. The models are trained on approximately six years of data (2012–2016) and validated using about one year of data (2017–2018). We compare the TPA-LSTM and Transformer models using (Formula presented.) 0.8 MeV electron fluxes and find that while the Transformer model performs slightly better, the difference is not statistically significant. Considering the Transformer's higher computational cost, we use the TPA-LSTM model to develop prediction models for electron fluxes of 275, 475, (Formula presented.) 0.8 MeV, and (Formula presented.) 2 MeV with a 5-min resolution at GEO, up to 3 days. The prediction efficiencies (PE) for 275, 475, (Formula presented.) 0.8 and (Formula presented.) 2 MeV electron fluxes based on about one year of test data (2018–2019) are 0.799, 0.831, 0.849, 0.881 (1-day prediction); and 0.551, 0.618, 0.663 and 0.710 (3-day prediction), respectively. Our high-time-resolution persistent models should be useful for both protecting satellites at GEO and serving as boundary conditions for physics-based radiation belt models.