[WIP] An automated ML pipeline to forecast hourly electricity demand for the PJM balancing authority.

Goal

The goal of this project is to demonstrate:

• ML Ops
  • ML workflow orchestration.
  • Versioning:
    • Dataset versioning.
    • Model versioning and experiment tracking.
  • ETL Pipeline.
  • Reliability:
    • Pipeline performance visibility and alerting. (WIP)
    • Automated unit tests (WIP)
  • Isolation between development and deployed/production environment infrastructure.
  • Performance comparisons between (model, version)s and a non-ML baseline.
  • Hyperparameter tuning.
  • Online prediction service
• ML
  • Timeseries feature engineering.
  • Timeseries forecasting with XGBoost.
  • Timeseries cross validation.

Data

• Electricity demand timeseries: EIA Open Data's hourly electricity demand data.
• Weather data: Open-meteo
• Holiday calendar: Calendarific

Stack

• Dev Env: Docker Compose
• ML Workflow Orchestration: Prefect
• Experiment tracking: mlflow
• Model registry: mlflow
• Dataset version tracking and 'Data Warehouse': DVC (+ git repo)
• Runtime data validation: Great Expectations
• Model: XGBoost
• Prod Env: AWS Copilot-managed containers on Fargate.

Development

Prefect Flow deployment:

 prefect deploy --name DEPLOYMENT_NAME --prefect-file flows/deployments/FLOW_DEPLOYMENT_CONFIG.yaml

References

• Géron, Aurélien. Hands-on machine learning with Scikit-Learn, Keras, and TensorFlow. "O'Reilly Media, Inc.", 2022.
• Rob Mulla's Kaggle tutorial on timeseries forecasting with XGBoost
• Prefect ECS Workers