SolarFlare-Nowcast v2

Deep Learning-Based Short-Term Solar Flare Forecasting Using GOES X-Ray Observations

Solar flares are among the most energetic phenomena in the Solar System and can significantly impact satellite operations, radio communications, navigation systems, and power-grid infrastructure. Accurate short-term forecasting of major solar flares is therefore a critical challenge in operational space weather prediction.

SolarFlare-Nowcast v2 is a deep learning framework designed to predict the probability of an M-class or higher solar flare occurring within the next 60 minutes using historical GOES X-Ray Sensor (XRS) observations.

The project combines domain-inspired feature engineering with a hybrid CNN-LSTM architecture to capture both short-term precursor signatures and long-term temporal evolution patterns in solar X-ray flux measurements.

---

Project Highlights

• Forecasts M-class and X-class solar flares

• Uses 10 years of GOES-15 observations (2010–2020)

• Hybrid Convolutional Neural Network + LSTM architecture

• Handles severe class imbalance using Focal Loss

• Evaluated using operational forecasting metrics:

  • True Skill Statistic (TSS)
  • Heidke Skill Score (HSS)
  • Probability of Detection (POD)
  • False Alarm Ratio (FAR)

• Supports:

  • Model training
  • Real-time NOAA GOES data ingestion
  • Command-line inference
  • REST API deployment
  • Interactive Streamlit dashboard

---

Scientific Objective

Given the previous 6 hours (360 minutes) of GOES X-ray flux observations:

Predict whether an M-class or stronger solar flare will occur within the next 60 minutes.

This problem is formulated as a binary classification task under highly imbalanced conditions where flare events are rare relative to quiet-Sun periods.

---

Dataset

Data Sources

GOES X-Ray Flux Data

• NOAA National Centers for Environmental Information (NCEI)
• NOAA Space Weather Prediction Center (SWPC)

Solar Flare Event Catalog

• LMSAL Heliophysics Event Knowledgebase (HEK)

---

Training Period

Parameter	Value
Satellite	GOES-15
Time Span	2010–2020
Sampling Rate	1 minute
Input Window	360 minutes
Forecast Horizon	60 minutes

---

Feature Engineering

Five physically motivated channels are generated from raw GOES XRS measurements.

Feature	Description
xrs_short	log10(0.05–0.4 nm flux)
xrs_long	log10(0.1–0.8 nm flux)
xrs_ratio	Spectral hardness proxy
deriv_short	First-order temporal derivative
rolling_max	30-minute rolling maximum

These features were selected to capture early precursor behavior commonly observed before major flare events.

---

Model Architecture

CNN-LSTM Hybrid Network

Input Shape:

360 × 5

Pipeline:

Input Sequence

↓

1D Convolution Layers

↓

Batch Normalization

↓

ReLU Activation

↓

LSTM Encoder

↓

Fully Connected Layers

↓

Sigmoid Probability Output

Why CNN-LSTM?

The original implementation used a dense Multi-Layer Perceptron (MLP), which flattened temporal information and treated all measurements as independent features.

The upgraded CNN-LSTM architecture:

• Preserves temporal ordering
• Learns local flux gradients and precursor patterns
• Captures long-range temporal dependencies
• Provides improved representation learning for sequential solar activity data

---

Training Strategy

Component	Configuration
Loss Function	Focal Loss
Optimizer	AdamW
Scheduler	ReduceLROnPlateau
Mixed Precision	Automatic Mixed Precision (AMP)
Checkpoint Metric	Validation TSS

---

Evaluation Metrics

Traditional accuracy is not a meaningful metric for flare forecasting due to severe class imbalance.

A model predicting "No Flare" continuously may exceed 98% accuracy while providing zero operational value.

Instead, evaluation is based on:

Metric	Purpose
TSS	Primary operational forecasting metric
HSS	Skill relative to random chance
POD	Fraction of flares successfully detected
FAR	Fraction of issued warnings that were false alarms

The model selection criterion is maximum validation TSS.

---

Repository Structure

solarflare-nowcast/

├── train.py
├── infer.py
├── app.py
├── dashboard.py
├── Config.yaml
├── requirements.txt

├── checkpoints/
│   ├── best_model.pt
│   ├── last_model.pt
│   └── training_history.json

└── src/
    ├── dataset.py
    ├── focal_loss.py
    ├── metrics.py
    ├── model.py
    ├── predictor.py
    ├── goes_fetcher.py
    └── goes_loader.py

---

Installation

git clone https://github.com/karmakar-rahul/solarflare-nowcast.git

cd solarflare-nowcast

pip install -r requirements.txt

---

Model Training

Full Training:

python train.py

Smoke Test:

python train.py --smoke

Custom Configuration:

python train.py --config Config.yaml

---

Inference

Live NOAA Data

python infer.py --live

Local CSV

python infer.py --csv path/to/goes_data.csv

Custom Threshold

python infer.py --live --threshold 0.45

---

Interactive Dashboard

Launch Streamlit dashboard:

streamlit run dashboard.py

Features:

• Live GOES monitoring
• Real-time flare probability
• Historical flux visualization
• Operational warning system

---

REST API

Start FastAPI server:

uvicorn app:app --reload --port 8000

Endpoints:

GET /live
GET /health

API Documentation:

http://localhost:8000/docs

---

Hardware Requirements

Component	Minimum	Recommended
CPU	Modern x86 CPU	Intel i5 / Ryzen 5+
RAM	8 GB	16 GB
GPU	Optional	RTX 2050 (4 GB)+
Storage	5 GB	10 GB

---

Current Limitations

• Trained exclusively on GOES-15 observations
• No active-region magnetic field information
• No spatial localization of flare source regions
• Optimized primarily for M/X-class forecasting
• False alarms are expected due to recall-oriented optimization

---

Future Work

• Integration of SDO/HMI SHARP magnetic parameters
• Transformer-based sequence forecasting models
• Multi-horizon prediction (30 min, 1 hr, 3 hr, 6 hr)
• Probability calibration
• Ensemble forecasting systems
• Explainable AI methods for flare precursor identification

---

Research & Technical Contributions

This project demonstrates:

• Time-series forecasting
• Deep learning for scientific data
• Class-imbalance handling
• Operational model evaluation
• Scientific machine learning
• End-to-end ML deployment

---

Acknowledgements

This work was developed using the PARAM Utkarsh High Performance Computing (HPC) Infrastructure provided by CDAC India (Centre for Development of Advanced Computing).

The computational resources and HPC environment made large-scale model development, experimentation, and training possible.

Special thanks to:

• CDAC India
• PARAM Utkarsh HPC Facility
• NOAA NCEI
• NOAA SWPC
• LMSAL HEK

---

Author

Rahul Karmakar

M.Sc. Physics (Astrophysics) Assam University, Silchar

Research Interests:

• Space Weather Forecasting
• Solar Physics
• Machine Learning for Scientific Applications
• High Performance Computing
• Computational Astrophysics

---

License

This repository is intended for research, educational, and portfolio purposes.