This experimental, privately-made (no relation to Nordpool) machine learning model estimates the consumer price of electricity in Finland (VAT 25.5%) for about a week ahead. The model correlates predicted weather and electricity production with previous periods where these factors were similar.
The model is trained and the forecast is updated automatically a few times per day.
The calendar shows how the predicted cheap/expensive periods stack up by hour and by day.
Loading hourly prices…
| Length | Average ¢/kWh | Countdown | Starts |
|---|---|---|---|
| Loading data... | |||
Based on data and forecasts from FMI weather stations on the coast of the Gulf of Bothnia, with price data shown in the background. Wind over 2...3 gigawatts is favorable for the electricity price, and with wind below 1 gigawatt, the price can rise significantly if there are maintenance outages or cold temperatures at the same time.
The graph shows a 30-day history of previous forecasts. Major changes can occur, for example, if the weather forecast changes significantly, if there is an unexpected production outage, if the cause of the price disturbance is historically new or rare, or if the model has been recently updated. See the change log below.
The green line is the final Nordpool price. The blue lines are the forecast ensemble from previous days, before the Nordpool price was known.
This is a machine learning model that seeks correlations between different variables and learns what price each combination leads to based on history. This is not a time series forecast, but each hour gets its value independently based on this information:
Try rotating the UMAP chart below. Marker size shows how much each feature influences the predicted price. Points that sit near each other act similarly in the training data and provide reinforcing signals for the model. A negative correlation means the feature tends to push prices down when it increases.
The forecast only sees the factors listed above. The source code is open, and you are welcome to experiment with other features. Training data starts from the beginning of 2023.
Although the model can "predict" the past and historically familiar types of the future, price estimates in new and exceptional situations can be off. So do not, for example, adjust your house heating based on these numbers if you are not at home.
The model has been online since February 2024. The following changes have been made since then.
October 28, 2025:
Customised settings now persist automatically in the browser, data sources refresh themselves at least once every six hours, and visitor metrics are counted anonymously and without cookies via Simple Analytics (EU).
October 25, 2025:
Added a table that surfaces the forecast’s cheapest 3/6/12-hour windows within the selected day and hour span, so you could, for example, line up overnight EV charging for the lowest-cost hours in the coming days. Also introduced a 7-day hourly calendar view that highlights which hours are expected to be cheap or expensive.
October 16, 2025:
Added a UMAP feature map to the landing page. The chart explains which feature groups move together and how strongly each one influences the forecast.
October 13, 2025:
A HACS integration is now available for Home Assistant users to, for example, schedule electric car charging based on cheapest 6 hour period.
June 24, 2025:
Added a dropdown menu to the historical forecasts chart allowing users to limit the forecast display to 1-5 days ahead. This makes it easier to evaluate short-term forecast accuracy.
May 7, 2025:
Experimental volatility model/scaler now aims to identify hours with higher risk of strong price fluctuations due to low wind power forecast. Updated the forecast chart to display hours with spike risk, where upwards scaling was applied. This significantly impacts the daily mean price forecasts on low-wind days.
April 25, 2025:
Line chart included in the forecast. Wind power chart now visualizes the possible correlation between wind power and price. History chart can now visualize 3 and 6 hour mean price values, and has a zoom control.
February 5, 2025:
Language model selection may now change dynamically based on model availability.
January 30, 2025:
The main page is now available in Finnish and English. A Markdown-formatted version can be found on GitHub in Finnish and in English. See also the bilingual JSON version. These may be useful, for example, for those building home automation systems. The DeepSeek API has recently been quite congested, and it doesn't always agree to generate texts. If this situation continues for a longer period, we'll revert to the GPT-4o model, although in that case we'll likely only provide a single language version.
January 18, 2025:
Wind power model tuned. Fingrid's continuously updated forecast provides the first 36 hours from the update time, and the rest are estimated from the weather data of the Baltic Sea region and Finland. The realization is now updated retrospectively. The language model now sees intra-day data and potential price spike risks. Language model changed to DeepSeek-V3.
January 12, 2025:
Fingrid, wind power capacity: data was missing from the interface for the beginning of the year. Open-Meteo, solar radiation: the provided forecast values (W/m²) were negative for the coming days, by half a million Watts. Checks added for such situations. Baltic Sea wind power areas added to the pricing model. Low wind abroad can raise the price in Finland, even if domestic production is strong.
January 4, 2025:
The forecast now extends 7 days ahead. Swedish and Estonian import connections are now handled separately in the model, not just as a sum. An outage in the Estonian connection typically lowers the price, while the Swedish one raises it.
December 30, 2024:
The model now estimates the impact of solar energy on the price. Wind power forecast uses XGBoost again.
December 25, 2024:
The history graph no longer shows old forecasts where the realized Nordpool price for tomorrow was already known to the model. History extended to 30 days.
December 24, 2024:
Information about holidays is now included in model training.
November 23, 2024:
Wind power forecast now experimenting with a neural network model (PyTorch).
November 17, 2024:
The language model now sees unplanned production outages of nuclear power plants. Language model changed to GPT-4o. The previous one was GPT-4o-mini. This should produce clearer Finnish.
November 10, 2024:
Fingrid no longer reports transmission capacity forecasts. The new JAO data source used in flow-based calculation only reports realized capacity. The forecast now assumes (again) that the latest realized transmission capacity remains for the forecast period.
October 20, 2024:
The language model now sees advance information on transmission connections and nuclear power plant maintenance.
October 17, 2024:
Reported maintenance outages and future production capacities of nuclear power plants (ENTSO-E) returned to the forecast. Price forecast changed to a bar chart.
September 22, 2024:
Planned transmission capacity (Fingrid) is now considered in the transmission forecast. Wind power data included in the language model's explanation. Wind power model tuned.
September 16, 2024:
Implemented a custom XGBoost wind power model based on historical and forecast data from 20 weather stations.
August 31, 2024:
Changed the forecast model type from Random Forest to XGBoost based on Optuna comparison.
August 19, 2024:
Added transmission capacities towards Finland: Northern Sweden, Central Sweden, and Estonia. This should refine the forecast during maintenance periods. The calculation assumes that the announced capacity for tomorrow remains the same for the 5-day forecast period.
August 11, 2024:
Removed ENTSO-E REMIT UMM production variable because according to the interface, OL3 would be under maintenance until 2025. Fortunately, this is not true, but it disrupts the forecast. We are investigating the issue and will return the nuclear power forecast to training if the cause of the error is found. The nuclear power forecast now assumes that the latest known production figure remains for the next 5 days.
Monthly data (January-December) has been removed from model training because the monthly averages of previous years can be very different from this year due to market conditions. The seasonal price effect is also evident in training from the weather when temperatures drop below freezing and consumption increases.