Predicting Short-Term X-ray Flux with Deep Learning Model

Predicting soft X-ray flux enhancements is crucial for early warning of solar storm impacts on Earth’s space environment and technological systems. In this work, we retrieve one-minute averaged science-quality data from NOAA Space Weather Prediction Center (SWPC) covering the period from 31 March 2010 to 4 March 2020, and construct two datasets: Dataset-A for analyzing the impact of look-back window length, and Dataset-B for five-fold cross-validation to compare model predictive performance and robustness. We apply four deep-learning models, iTransformer, BiLSTM, BiLSTM-CNN, and Seq2Seq, along with MLP baseline, to forecast short-term X-ray flux. Furthermore, we develop a real-time operational system for forecasting X-ray flux using the iTransformer model. The main results are as follows: (1) Based on Dataset-A, we conduct extensive experiments to investigate the impact of different look-back window sizes on soft X-ray flux prediction performance. The results demonstrate that the size of the look-back window does not significantly affect the performance of any model. (2) In the five-fold cross-validation, the mean and variance of the MAE for iTransformer are better than those of the other models. (3) For predictions at future horizons of 1, 5, 10, 15, 20, 25, 30, 35, and 40 min, the mean MAE of iTransformer is 1.36, 1.74, 2.03, 2.24, 2.49, 2.70, 2.93, 3.10, and 3.28 × 10−7, respectively. To the best of our knowledge, this is the first study to perform short-term soft X-ray flux forecasting and to apply the iTransformer model for this purpose.