SEP-Prediction

Solar Energetic Particle Forecasting with Multi-Task Deep Learning: SEPNet

Overview

SEP-Prediction is an open-source framework for forecasting Solar Energetic Particle (SEP) events using advanced multi-task deep learning approaches. This project leverages comprehensive space weather data, including solar flares, coronal mass ejections (CMEs), and solar magnetic field parameters (SHARP), to deliver reliable SEP predictions and alerts. Our main model, SEPNet, combines LSTM and transformer architectures to capture both temporal trends and contextual dependencies in solar activity.

Features

• Joint prediction of SEP occurrence and solar activity
• Incorporation of flare, CME features, and SHARP magnetic field parameters
• Benchmarking deep learning methods against classical machine learning algorithms
• Ready-to-use datasets and scripts for easy replication and further experimentation

Repository Structure

• data/ — Contains preprocessed datasets for model training and evaluation:
  • df_CME.csv, df_flare.csv, df_SEP.csv: Collected event data
  • df_sharp.csv, df_sharp_imputed.csv: Large SHARP magnetic parameter files (accessible via Google Drive link)
  • rolling_combined.csv: Main preprocessed feature set using rolling window sampling (constructed via Data-Construct.R)
  • rolling_combined_training.csv, rolling_combined_testing.csv: Train/test splits based on SEPVAL event lists
  • SEPValidationChallengePhaseIII_NonEvents_v4.csv, SEPValidationChallengePhaseIII_SEPevents_v5.csv: SEPVAL event definitions (more info)
• data_nrt/ — Near real-time data files for operational testing
• data_unprocessed/ — Unprocessed raw data (part) and external sources:
  • CME data from DONKI
  • Flare data with SunPy instructions
  • SHARP magnetic data info: JSOC documentation
• Data-Construct.R & Data-Construct-nrt.R — Data processing and feature engineering scripts
• multitask_model.py — SEPNet multi-task deep learning model implementation
• Main_SEPVAL.ipynb — Main notebook for SEPNet and SEPNet2 experiments with SEPVAL dataset
• OtherGeneralModels_SEPVAL.ipynb — Classical ML baselines (logistic regression, random forest, SVM, XGB)
• Main_Stratified.ipynb — Stratified experimentation notebook, targeting classification of general SEP (>10 MeV 1e-6 pfu)
• Main_Stratified_Operational_SEP.ipynb — Stratified experimentation notebook, targeting classification of operational SEP (>10 MeV 10 pfu)
• Main_RealTime.ipynb — Notebook for real-time SEP event prediction

Contact

For questions or collaboration, contact yuyian@umich.edu.

1. Install required packages:

• We modeled with python 3.12 and R 4.5.1
• See requirements listed in each notebook (.ipynb) and R (.R)

2. Access large files:

• Download df_sharp.csv and df_sharp_imputed.csv from Google Drive.

2. Run notebooks:

• Use the Jupyter notebooks for experiments, evaluation, and model training.

Data Availability

• All code and most datasets are included directly in this repository.
• Large magnetic field parameter files (SHARP) are provided via a dedicated Google Drive link.

Citation

If you use SEP-Prediction, please cite our associated publication (to be linked).