Skip to content

What is EPF?

EPF (Electricity Price Forecasting) is a machine learning system that predicts Spanish day-ahead electricity market (OMIE) prices up to 7 days ahead, at both hourly and 15-minute resolution.

Two Forecast Products

EPF delivers two operationally distinct forecasts, aligned with OMIE market publication times:

  • D+1 Day-Ahead — 24 hourly prices for tomorrow, delivered before OMIE gate closure at ~10:00 UTC
  • D+2-D+7 Strategic — 144 hours of forward-looking prices, delivered at ~15:00 UTC using published D+1 actuals as model input

Key Capabilities

  • 7-day horizon with hourly and 15-minute resolution
  • LSTM-XGBoost hybrid (day-ahead): task-aligned LSTM encoder + 90 tabular features
  • Gradient boosting ensemble (strategic): HistGradientBoosting, LightGBM, XGBoost averaged
  • Confidence intervals via split conformal prediction (50% and 90% bands)
  • 150+ features spanning price dynamics, temporal embeddings, weather, generation mix, and commodities
  • Walk-forward backtesting with comprehensive evaluation metrics

Data Sources

The system ingests data from four sources:

SourceData
REE/ESIOS16 electricity system indicators (demand, generation, interconnections, prices)
Open-MeteoHourly weather from 5 Spanish stations (population-weighted)
TTF GasNatural gas prices (EUR/MWh)
EU ETSCarbon permit prices (EUR/tCO2)

Current Performance

ProductModelMAEBiasWindow
D+1 Day-AheadLSTM-XGBoost hybrid (v10.1)15.73 EUR/MWh-0.65 EUR/MWh150-day (incl. crisis)
D+2–D+7 StrategicEnsemble (v4.3)19.79 EUR/MWh-0.30 EUR/MWh149-day

The day-ahead bias of -0.65 EUR/MWh is the best ever recorded — 19× better than the previous -12 EUR/MWh. The 150-day v10.1 window includes the March 2026 Iran crisis (prices 170–247 EUR/MWh), adding ~1–2 EUR/MWh to MAE versus crisis-free periods. v4.3 remains in production for strategic horizons.

See the methodology changelog for the full experiment history.