Day-ahead electricity load forecasting
A complete forecasting pipeline for Dutch electricity demand, from ENTSO-E and weather ingestion to calibrated prediction intervals, scheduled inference, and live scoring.
01 / The problem
Electricity demand is easy to forecast retrospectively and much harder to forecast honestly. Features that look useful can be unavailable at the real forecast origin, and a point prediction hides the uncertainty that operators actually need.
02 / The approach
The pipeline freezes chronological splits, fits transformations on training data only, removes short lags that would leak information across a full next-day horizon, and calibrates quantile predictions with Conformalized Quantile Regression. Training and live inference share the same feature-construction path.
03 / What shipped
- Leakage-checked feature engineering with explicit forecast-origin rules
- Quantile forecasts calibrated on validation and verified on a held-out test split
- Daily ingestion, inference, scoring, and storage checks through GitHub Actions and Supabase
- A documented path from historical weather to real NWP forecasts
04 / Outcome
The resulting system produces a point forecast and a calibrated interval, then runs daily ingestion, prediction, scoring, and sanity checks through scheduled workflows. The README documents the idealized-weather limitation instead of presenting it as an operationally solved problem.