Flare Transformer

• An overview of the Flare Transformer.

1. Requirements

• Python 3.6.9
• Ubuntu 18.04

2. Installation

• In the following procedure, ~/work is assumed to be used as a working directory.

$ cd ~/work
$ git clone URL
$ cd flare_transformer
$ pip install -U pip
$ pip install -r require.txt

3. Download data

• The required data files should be put under data/ folder.
• Visit http://wdc.nict.go.jp/IONO/wdc/solarflare/index.html and download the database of physical features.

$ cd ~/work/flare_transformer
$ mv ~/data.zip data/
$ unzip data/data.zip

• Visit https://sdo.gsfc.nasa.gov/data/ and download hourly magnetograms.

$ mv ~/magnetogram_images.tar.gz data/
$ tar -zxvf data/magnetogram_images.tar.gz

4. Preprocess

• Preprocess the physical features and magnetogram images by the following procedure.

$ python src/preprocess.py

• The following data files should be created under data/.
  • data/data_20XX_magnetogram.npy
  • data/data_20XX_feat.csv
  • data/data_20XX_window.csv
  • data/data_20XX_label.csv

5. Training

$ cd ~/work/flare_transformer
$ ./train.sh

• Training example is shown in train.sh. --params should be specified according to your settings.
• --params takes a path of a JSON file. In the JSON file, values for parameters such as learning rate, batch size, etc. should be specified.
• A sample of a JSON file is shown in params/params2017.json.

6. Results

• Quantitative Results

  • We report the model performance of the Flare Transfomer as follows:

    	GMGS
    Flare Transformer	0.503 ± 0.059	0.530 ± 0.112	0.082 ± 0.974

  • We also report the confusion matrix for the 2017 test set as follows:

  

• Qualitative Results

  • The figure below shows line-of-sight magnetograms from September 3th, 2017 23:00 UT to September 5th, 2017, 23:00 UT. An X-class solar flare occurred at 12:02 on September 6, 2017, and the model was able to predict the correct maximum solar flare class.