End-to-end ML pipeline identifying high-risk electric utilities across the United States using EIA-861 reliability data and NOAA Storm Events data.
🔗 Live Dashboard → sejjjallll-power-outage-risk-dashboard.hf.space
| Metric | Value |
|---|---|
| Records processed | 3,441,222 raw → 1,677 utility-level |
| Features engineered | 46 |
| Utilities classified High Risk | 336 of 1,677 (20%) across 50 states |
| MLflow experiments | 28 (7 classifiers × 2 feature sets × 2 thresholds) |
| Best model | Logistic Regression + Weather features |
| ROC-AUC | 0.668 (leakage-corrected from 1.0) |
| PR-AUC | 0.325 |
| Weather feature uplift | +54% ROC-AUC (0.50 → 0.77) |
| Estimated economic loss | $74.5B annually |
Live Dashboard (Gradio + Folium on Hugging Face Spaces)
FastAPI Endpoint (Swagger UI)
FastAPI Prediction Response
MLflow — 28 Tracked Experiments
Stage 1 — src/preprocess.py Raw CSV (3.44M rows) -> Clean Parquet
Stage 2 — src/features.py Clean Parquet -> 1,677 utility-level features
Stage 3 — src/train.py 28 experiments -> best_model.pkl + MLflow logs
Stage 4 — api/main.py FastAPI REST endpoint with /predict /health /docs
Dashboard — dashboard/app_gradio.py Gradio + Folium maps deployed to HF Spaces
Initial models returned ROC-AUC = 1.0 — a clear signal of leakage.
Root cause: Three engineered features (estimated_annual_loss_usd, nerc_sla_breach_risk, sla_breach_margin_min) were derived directly from SAIDI, which defines the target variable. The model was predicting risk from risk.
Fix: Removed all SAIDI-derived features from the feature set, retrained across all 28 experiments. ROC-AUC corrected from 1.0 → 0.668.
This is the difference between a model that looks good in development and one that would actually work in production.
ML/Data: Python · Scikit-learn · XGBoost · LightGBM · MLflow · Pandas · NumPy API: FastAPI · Pydantic · Uvicorn Dashboard: Gradio · Folium · Plotly DevOps: Docker · GitHub Actions (CI/CD) · Hugging Face Spaces
git clone https://github.com/SejalKhade/Power-Outage-Risk-Dashboard.git
cd Power-Outage-Risk-Dashboard
pip install -r requirements.txt
# Run pipeline stages
python -m src.preprocess # 3.44M rows -> clean parquet
python -m src.features # -> 1,677 utility features
python -m src.train # 28 MLflow experiments
# Run API
uvicorn api.main:app --reload # http://localhost:8000/docs
# Run dashboard locally
python dashboard/app_gradio.py # http://localhost:7860
# Or run containerized API
docker build -t power-outage-risk .
docker run -p 8000:8000 power-outage-riskpower-outage-risk-dashboard/
├── src/
│ ├── preprocess.py Stage 1 — data pipeline
│ ├── features.py Stage 2 — feature engineering
│ ├── train.py Stage 3 — ML training + MLflow
│ └── utils.py Shared helpers
├── api/
│ └── main.py FastAPI REST endpoint
├── dashboard/
│ └── app_gradio.py Gradio dashboard (HF Spaces)
├── outputs/models/ best_model.pkl, sensitivity_results.csv, metrics.json
├── docs/images/ Screenshots
├── .github/workflows/ci.yml GitHub Actions CI pipeline
├── Dockerfile
├── requirements.txt
└── README.md
Sejal Khade MS Data Science · University of Texas at Arlington · May 2026