Return
Multiclass solar flare forecasting models with different deep learning algorithms
DOI:10.1093/mnras/stad839.png)
Abstract
En 中文
We develop a Hybrid Bidirectional Long and Short-Term Memory based on attention mechanism (HBiLSTM-Attention) model and a BiLSTM-Attention model for multiclass flare forecasting within 24 h. We construct a new data base containing 10 separate data sets with magnetogram images and magnetic field parameters. Based on the same data base, for the first time we compare the multiclass forecasting performance of our proposed HBiLSTM-Attention model, BiLSTM-Attention model, and three other deep-learning models based on Convolutional Neural Network (CNN-based) from two aspects of categorical performance with the true skill statistic (TSS) and probabilistic performance with the Brier skill score (BSS). The major results are as follows. (1) The TSS values of our proposed model are 0.692 +/- 0.042, 0.475 +/- 0.038, 0.642 +/- 0.043, 0.754 +/- 0.062, 0.692 +/- 0.042, and 0.708 +/- 0.052 for No-flare, C, M, X, >= C, and >= M class, respectively, which are better than those of the BiLSTM-Attention model, and much better than those of the three other CNN-based models. (2) Our proposed model achieves the scores of BSS = 0.498 +/- 0.061, 0.202 +/- 0.037, 0.209 +/- 0.050, -0.271 +/- 0.180, 0.498 +/- 0.061, 0.268 +/- 0.056 for No-flare, C, M, X, >= C, and >= M class, respectively, outperforming the other four models in every class except for X class. (3) To our knowledge, HBiLSTM-Attention is the first multiclass flare forecasting model based on magnetic field parameters and deep learning, and achieves promising prediction performance. Moreover, this is the first attempt to investigate the reliability of probabilistic prediction for multiclass flares.
Keywords:
methods: data analysis
techniques: image processing
Sun: activity
Sun: flares
Sun: magnetic fields
Journal
IF:
4.8
Papers:
7.0W
Citations:
25.0W
Organization
Cited Papers
“Worse than I anticipated” or “This isn’t so bad”?: The impact of affective forecasting accuracy on self-reported task performance
PLOS ONE
IF0

