<<<<<<< HEAD solar_wind.csv geomagnetic_model.pkl These files are too big for Github, so please download it from here.

To run the app, please run python app.py in the terminal

Geomagnetic Storm Predictor

A small Flask workbench for estimating the Dst geomagnetic storm index from hourly solar-wind conditions. It supports one-off predictions, realistic scenario fill-ins, CSV batch runs, and storm severity labels.

Setup

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Train the model

The trained model file (geomagnetic_model.pkl) is large and not checked in. Generate it from the raw data:

python train_model.py

If the merged dataset (hourly_avg_dst.csv) is missing, train_model.py rebuilds it from the raw solar_wind.csv + labels.csv via prepare_data.py, then trains and writes geomagnetic_model.pkl. You can also run the data step on its own:

python prepare_data.py   # solar_wind.csv + labels.csv -> hourly_avg[_dst].csv

Note: timedelta restarts per period (train_a/b/c are separate time series), so the data is aggregated by (period, timedelta) rather than by timedelta alone.

Run the app

python app.py

Then open http://127.0.0.1:5000.

The app expects these five inputs:

speed,bt,temperature,bz_gsm,density
420,8,90000,-4,6

CSV batch uploads use the same column names. The batch response reports how many rows were predicted and how many were skipped because of missing or invalid data.

Model performance

A full technical report — background, data, methodology, metric definitions, model comparisons (Logistic Regression / SVM / Random Forest / XGBoost), operating thresholds, calibration, and all graphs — is in RESULTS.md.

Headline numbers (honest, time-aware evaluation of the temporal model):

• Forward-chaining split: R² ≈ 0.67, MAE ≈ 6.7 nT, storm AUC ≈ 0.98, AP ≈ 0.71
• Strictest unseen-epoch (period-holdout) ceiling: R² ≈ 0.59, AP ≈ 0.63
• A random split inflates these (R² 0.44 / AP 0.48); temporal features are the biggest accuracy lever, and partitioning by regime gave no benefit. Full analysis and graphs in RESULTS.md.

Regenerate everything:

# Initial analysis (random split)
python evaluate_model.py        # regression + classification diagnostics
python compare_models.py        # LogReg vs SVM vs Random Forest
python compare_models_tuned.py  # + tuning, class balancing, XGBoost
python tune_and_finalize.py     # XGBoost tuning, thresholds, calibration

# Leakage correction & time-aware modeling
python prepare_features.py      # temporal + sunspot + propagation features
python eval_splits.py           # random vs period-holdout vs forward-chaining
python eval_temporal.py         # base vs temporal vs +auxiliary features
python eval_tail.py             # extreme-storm tail remedies

Data

Raw inputs (from the DrivenData MagNet dataset), too large for GitHub:

• solar_wind.csv — minute-cadence solar-wind measurements (the main model input)
• labels.csv — hourly Dst target
• satellite_pos.csv, sunspots.csv — additional context, not yet used by the model

prepare_data.py derives hourly_avg.csv and hourly_avg_dst.csv from these.

• solar_wind.csv: https://drive.google.com/file/d/1kb1oa8q9HKvXOLotmwH7iLxvDuzzsJyc/view?usp=sharing

0fa6bfe (initial)