1 of 7

Team: What’s new?

Yuan-Tung Chou,

Guo-Chi Li,

Tsung-Wei Huang,

Ting-Ju Wei,

Tzu-Yao Chien,

Po-Chih Kuo

Save the Earth from Another Carrington Event!

2 of 7

2

How to Predict Solar Storm Event?

Massive Geomagnetic Storm: Coronal Mass Ejection From the Sun Could Knock Out the Power Grid and Internet

Solar storm affects your life, it could disrupt

  • the electric power grid
  • satellite operations
  • radio systems

​

+

+

+

+

+

from NOAA website

ICON by winnievinzence

wind icon by Freepik

Data from DSCOVR satellite (DSCOVR-mag)

ML Models

Solar Winds

Event Warnings

predict

model training

?%

3 of 7

3

Problem Formulation

  • Instead of directly using DSCOVR magnetic data to predict storm probability, the problem is divided into 3 parts (3 AI models):

DSCOVR magnetic data

AI model

Event Probability

noise, distortion

➝ less representative

more related to storm but not used

➝ missing important features

proton data

DSCOVR magnetic data

Wind magnetic data

map

Wind

proton data

fit

Event Probability

predict

cleaner data

more related to storm event

AI

AI

AI

each of the AI model can be used in downstream tasks

each of the AI model are more representative

4 of 7

4

Model Architecture

  • Data is preprocessed before entering the models
    • The 3 models (trained separately):

preprocessed

X

Y

DSCOVR-mag

RNN(Seq2Seq)

Wind-mag

1. Regression Problem

pred-Wind-mag

RNN(Seq2Seq)

Wind-proton

2. Regression Problem

DSCOVR-mag

RNN(Seq2Val)

Storm Event Prob

3.

Logistic

Regression Problem

Extra Dataset:

DST on Earth

https://wdc.kugi.kyoto-u.ac.jp/dstdir/index.html

5 of 7

5

Storm Event Forecasting Results

  • Model 3 detects all of the event in the below case.
  • Model 3 tends to be conservative.

testing dataset

  • accuracy = 0.9551
  • recall = 1.0000
  • precision = 0.4478

6 of 7

6

Model 1 Result

DSCOVR-Wind R2 = 0.551

Prediction-Wind R2 = 0.978

  • Model 1 successfully maps DSCOVR data to Wind’s.
  • Model 1 fixes the noise and distortion in DSCOVR data.

Time (hr)

7 of 7

7

Insights and Summary

  • The proposed 3-part training pipeline works well on the dataset.
  • It can forecast the probability solar storm in real time (It only takes 0.46ms to predict one month’s storm probability).
  • Furthermore, each of the model can not only be used in forecasting, but also in other future tasks.
  • Examples are shown that model can detect upcoming solar storms.