17:15 〜 18:45
[PEM11-P05] Solar flare prediction modeling using feature compression with convolutional autoencoder
キーワード：太陽フレア
The recent improvement in machine learning methods enables us to predict solar flares more accurately. Sun et al. (2022) developed the machine learning model whose prediction probability is calculated as a weighted linear sum of the CNN model for magnetograms and the LSTM model for the time series of the SHARP parameters. The model achieved ACC>=0.91, which indicates that it is critical for the flare prediction to use both spatial and time series features of the active region. However, previous studies have trained CNN and LSTM separately, and it is not obvious whether the model is optimized for the mixture of spatial and time series information. To this end, we address the improvement of flare prediction accuracy by optimizing for spatial and time series information simultaneously using the feature compression method.
We used approximately 140,000 magnetograms at 3-hour intervals from the SHARP dataset from May 2010 to December 2017 observed by SDO/HMI. We use the SWAN dataset as label data and define magnetograms as “flared” where M-class flare or greater occurred within 24 hours. The training and validation data do not contain the same active regions as those in the test data. Our model construction consists of three steps. In the first step, we train the convolutional autoencoder with magnetograms as a compression model and train LSTM part by using time series data of 9 magnetograms, e.g. 24 hours in total, in the second step. In the last step, we train the whole model, autoencoder and LSTM parts, from the trained parameters obtained in the first and second steps.
As a result, we achieved TSS=0.820 for M-class flare or greater. In contrast, we obtained TSS=0.647 without the first and second steps, which suggests that feature compression and step-by-step learning are effective in learning the solar flare occurrence with limited observational results.
We used approximately 140,000 magnetograms at 3-hour intervals from the SHARP dataset from May 2010 to December 2017 observed by SDO/HMI. We use the SWAN dataset as label data and define magnetograms as “flared” where M-class flare or greater occurred within 24 hours. The training and validation data do not contain the same active regions as those in the test data. Our model construction consists of three steps. In the first step, we train the convolutional autoencoder with magnetograms as a compression model and train LSTM part by using time series data of 9 magnetograms, e.g. 24 hours in total, in the second step. In the last step, we train the whole model, autoencoder and LSTM parts, from the trained parameters obtained in the first and second steps.
As a result, we achieved TSS=0.820 for M-class flare or greater. In contrast, we obtained TSS=0.647 without the first and second steps, which suggests that feature compression and step-by-step learning are effective in learning the solar flare occurrence with limited observational results.