eupowerprices.com

Methodology

How the forecast works, what the models use, and the main limitations of the approach.

How the Forecast Works

The forecasts on this site are generated using machine learning models trained on historical electricity-market, power-system, and weather data. The forecast extends up to 14 days ahead and is refreshed several times each day.

Two-stage architecture

First, dedicated models forecast wind generation, solar generation, and electricity demand. Depending on the model, they use weather forecasts, recent observations, calendar information, and other relevant inputs. Their outputs are then passed to a downstream price model together with market and power-system features.

The price model uses gradient-boosted decision trees to produce an hourly prediction for each forecast period from the inputs available for that hour.

Training and weather data

Historical training data currently extends back to 2023. Weather inputs use ECMWF forecasts delivered through Open-Meteo, including archived forecasts for model training, together with market data from ENTSO-E and other publicly available sources.

Model Inputs

The downstream price model combines several types of information:

  • Forecast fundamentals: electricity demand, wind generation, and solar generation.
  • Power-system conditions: regional nuclear availability, hydrological conditions, recent hydro generation, and forecast cross-zonal transfer capacity.
  • Regional context: relevant neighboring prices and wind conditions in connected markets.
  • Temporal and market context: hour, weekday, season, public holidays, direct weather summaries, and recent price patterns.

What the Model Learns

The model identifies patterns that have historically been associated with electricity prices. For example, higher renewable generation generally puts downward pressure on prices, while stronger demand and tighter supply conditions tend to support prices. Weather can also affect prices indirectly through demand and generation availability.

Rather than explicitly simulating the physical operation of the power system, the model learns these statistical relationships from historical data.

Forecast Uncertainty

The central forecast is a point estimate rather than a guarantee. Where available, the Weather Ensemble page runs the same trained models across alternative ECMWF weather scenarios. This illustrates uncertainty arising from the weather forecast, but it does not include uncertainty from outages, market behavior, fuel prices, data errors, model error, or other unmodeled developments.

Geographic Coverage

Forecasts are currently available for:

  • Austria (AT)
  • Belgium (BE)
  • Czechia (CZ)
  • Denmark (DK1, DK2)
  • Estonia (EE)
  • Finland (FI)
  • France (FR)
  • Germany (DE)
  • Latvia (LV)
  • Lithuania (LT)
  • Netherlands (NL)
  • Norway (NO1-NO5)
  • Poland (PL)
  • Portugal (PT)
  • Slovakia (SK)
  • Spain (ES)
  • Sweden (SE1-SE4)

Limitations

Statistical rather than physical modeling

The approach is statistical rather than a full market simulation. It does not explicitly optimize generation dispatch, reservoir operation, transmission flows, bids, or unit commitment. Instead, it learns relationships present in historical market and power-system data. Factors that are unavailable or not explicitly modeled, including fuel and emissions markets, unplanned thermal outages, bilateral trading, and participant behavior, can therefore cause actual prices to differ materially from the forecast.

Transmission

Cross-zonal transmission is represented using forecast transfer capacities published by ENTSO-E. This captures changes such as planned maintenance, but it does not reproduce the full market-coupling and network-constraint calculation used in actual market clearing. Congestion and cross-border effects are therefore approximated rather than physically simulated.

Hydro treatment

Hydro generation in Finland, Norway, and Sweden is currently represented using the latest observed daily generation level carried forward through the forecast horizon. A previously tested statistical hydro forecast did not consistently outperform this persistence baseline, so the simpler approach is currently used.

Hydrological state variables such as reservoir levels, river flow, snow, and precipitation remain separate price-model inputs. The approach does not simulate reservoir optimization or hourly hydro dispatch, so it cannot anticipate changes in operating strategy or genuine shifts in hydro generation. Its reliability therefore decreases with forecast horizon.

Unfamiliar and extreme conditions

Gradient-boosted trees are strongest when future conditions resemble combinations represented in their training data. They do not naturally extrapolate structural relationships far beyond the historical range. Extreme price spikes, negative-price events, and other unusual market situations can therefore be difficult to predict, particularly when similar events are rare in the historical record.

Structural market change

The electricity system changes over time as generation capacity, demand, transmission, and market rules evolve. Models are retrained and inputs such as installed capacity are updated, but historical relationships may still become less representative as the system changes.

Data Sources

The live forecasting system currently uses the following public and third-party data sources. The forecasts, derived features, and visualizations on this site are our own calculations.