A Proton Flux Prediction Method Based on an Attention Mechanism and Long Short-Term Memory Network

Abstract

Accurately predicting proton flux in the space radiation environment is crucial for satellite in-orbit management and space science research. This paper proposes a proton flux prediction method based on a hybrid neural network. This method is a predictive approach for measuring proton flux profiles via a satellite during its operation, including crossings through the SAA region. In the data preprocessing stage, a moving average wavelet transform was employed to retain the trend information of the original data and perform noise reduction. For the model design, the TPA-LSTM model was introduced, which combines the Temporal Pattern Attention mechanism with a Long Short-Term Memory network (LSTM). The model was trained and validated using 4,174,202 proton flux data points over a span of 12 months. The experimental results indicate that the prediction accuracy of the TPA-LSTM model is higher than that of the AP-8 model, with a logarithmic root mean square error (logRMSE) of 3.71 between predicted and actual values. In particular, an improved accuracy was observed when predicting values within the South Atlantic Anomaly (SAA) region, with a logRMSE of 3.09.