Jingjing Wang1,2, Bingxian Luo1,2,3, & Siqing Liu1,2,3
1 State Key Laboratory of Space Weather, National Space Science Center, Chinese Academy of Sciences, Beijing 100190, China
2 Key Laboratory of Science and Technology on Environmental Space Situation Awareness, Chinese Academy of Sciences, Beijing 100190, China
3 University of Chinese Academy of Sciences, Beijing 100190, China
Solar flare is one of the major eruptive phenomena that occur in solar atmosphere. Developing solar flare prediction is one of the essential methods to prevent the possible storm-like geo-effectiveness. Space Weather HMI Active Region Patches (SHARP), which consists of vector magnetic field images and many other magnetic properties of ARs, has been widely used by the community. Many researchers have investigated the datasets, then developed and applied solar flare prediction models using them. In recent years, machine-learning algorithms have been applied in many such investigations and progress has been made. Considering the needs of solar flare forecast in the operational application at space weather prediction centers, it is important to develop a model that can provide quantitative time-in-advance with promising evaluation metrics (i.e., prediction accuracy).
In Ref , an anomaly detection algorithm has been adopted to identify the precursor for strong flares, and a strong flare classification model was developed. We assume that the magnetic configuration of active regions (ARs) in quiet periods has certain similarity and can be considered as “normal” features, while there are some other magnetic features of active regions that are related to strong flares. They can be considered as precursors of strong flares and “anomaly” features. This study is aiming to identify those “anomalies” and apply them in strong flares’ forecasting.
Figure 1| Radial magnetic field of AR12192 (corresponding to SHARP patch number 4698) before X1.6 flare erupted at 14:28UTC on 2014 October 22. There are ten slices (with a size of 320 × 320 pixels) cut randomly from the original observations.
As shown in Figure 1, we first cut randomly the SHARP Br patches into several square slices (with sizes of 320 × 320 pixels) to obtain a uniform and standard dataset. For fast computation purpose, we compress the slices into smaller ones (with a size of 32 × 32 pixels) as shown in the first and third rows in Figure 2. The SHARP Br dataset of non-strong-flare ARs in 2010-2014 has been used to train an unsupervised auto-encoder network. Taken an original picture as input, the network will produce a reconstrued picture. The second and fourth rows in Figure 2 are the reconstructed pictures corresponding to the original pictures in the first and third rows. Those similar features of non-strong-flare ARs can be understood and memorized by the auto-encoder network.
Figure 2| Radial magnetic field of AR12192 (corresponding to SHARP patch number 4698) before X1.6 flare erupted at 14:28UTC on 2014 October 22. There are ten slices (with a size of 320 × 320 pixels) cut randomly from the original observations.
We investigate the mean standard errors (MSEs) of the original pictures and reconstruct pictures derived by the auto-encoder network, and conduct a statistical analysis of the non-strong-flare samples. Then we carry on the precursor identification for strong flares based on anomaly detection algorithm and develop a strong-flare classification model. Applying the classification model on the non-strong-flare ARs sample in 2015-2019 and strong-flare ARs sample in 2010-2019, we take the MSEs computed from the results by the auto-encoder network as a predictor, and compare them with a specific threshold, and finally get a strong-flare or non-strong-flare forecast for each sample. We find that the “anomaly” magnetic features of ARs that indicate the strong flares can be detected and used to predict strong flares. We then evaluate the performance of the classification model and obtain the quantitative time-in-advance for strong-flare prediction.
The strong-flare classification model reaches an F1 score of 0.8139. Moreover, for those correctly predicted strong-flare events (94 M-class flares and above), the model reaches an average first warning time of 45.24 hours. The results indicate that 1) the anomaly detection algorithm can be used in precursor identification for strong flares and help in both improving strong-flare prediction accuracy and enlarging the time in advance. 2) the precursor magnetic structures of ARs appear some time before the strong-flare eruption, and thus the average maximum warning period for strong-flare prediction is close to two days. And the obtained average maximum warning period for strong-flare prediction (nearly two days) will be useful for future applications for space-weather solar-flare prediction.
 Wang, J., Luo, B., & Liu, S. 2022, Front. Astron. Space Sci.,