German Power Price Forecasting
Abstract
Forecasting hourly day-ahead electricity prices in Germany (2020 to 2026) around the merit-order effect, from automated ETL to ARIMAX vs Ridge vs XGBoost.
- Methods
- ARIMAX · Ridge · XGBoost · Time-series features
- Stack
- Python · XGBoost · statsmodels · scikit-learn · SMARD API · Open-Meteo
When wind and solar are abundant, they displace expensive gas and coal at the margin and the price falls: the merit-order effect. The model is built around that mechanism.
Pipeline
- ETL from SMARD (prices, load, generation by fuel), Open-Meteo (hourly weather) and Yahoo Finance (TTF gas, EUA carbon, Rotterdam coal).
- Weather clusters: five strategic locations (Hamburg, Helgoland, Munich, Berlin, Cologne) as proxies for national generation and demand.
- Features: 24 h and 168 h lags for daily and weekly cycles, net load.
Models
A Ridge baseline for fundamental trends, ARIMAX(2,0,1) with net load and wind as exogenous regressors, and XGBoost to capture the step-function shape of the merit-order curve.