5:15 PM - 7:15 PM
[PEM10-P04] Development of a Solar Flare Prediction Model Using Feature Compression Method and Weighted Images
Keywords:Solar Flare
Solar flares are explosive phenomena in which the energy accumulated in the magnetic fields around sunspots is released. In research by Kubo et al. at the JpGU 2024 meeting, it was demonstrated that by using feature compression and conducting the learning process step by step, the CNN+LSTM model can be efficiently trained even with a limited amount of data. Based on the previous research by Kubo et al., this study aims to improve the prediction accuracy by efficiently learning the magnetic field structure near magnetic neutral lines using weighted images.
We used the SHARP dataset from SDO/HMI, covering the period from May 2010 to December 2017, with approximately 140,000 magnetic field images sampled at 3-hour intervals. The SWAN dataset was used for labeling, where cases with M-class or higher flares occurring within 24 hours were classified as "flare occurrence." The training and validation data were split based on time periods.
The model construction process is as follows. First, a Convolutional Autoencoder (CAE) was trained on preprocessed magnetic field images. Next, magnetic neutral lines were extracted by the signal processing method, and weighting was applied to the difference images based on the distance from these lines. Subsequently, a second CAE was trained on the weighted difference images. Following this, an LSTM was trained using the features extracted from both CAEs. Finally, the encoder parts of both CAEs and the LSTM were trained.
We obtained a high value of TSS=0.928, significantly surpassing the research by Kubo et al. (TSS=0.820) and other previous studies. This can be interpreted as the result of the weighting by magnetic neutral lines working effectively.
We used the SHARP dataset from SDO/HMI, covering the period from May 2010 to December 2017, with approximately 140,000 magnetic field images sampled at 3-hour intervals. The SWAN dataset was used for labeling, where cases with M-class or higher flares occurring within 24 hours were classified as "flare occurrence." The training and validation data were split based on time periods.
The model construction process is as follows. First, a Convolutional Autoencoder (CAE) was trained on preprocessed magnetic field images. Next, magnetic neutral lines were extracted by the signal processing method, and weighting was applied to the difference images based on the distance from these lines. Subsequently, a second CAE was trained on the weighted difference images. Following this, an LSTM was trained using the features extracted from both CAEs. Finally, the encoder parts of both CAEs and the LSTM were trained.
We obtained a high value of TSS=0.928, significantly surpassing the research by Kubo et al. (TSS=0.820) and other previous studies. This can be interpreted as the result of the weighting by magnetic neutral lines working effectively.