Spot Forecast

Day-ahead electricity prices for Germany, Luxembourg, the Netherlands, Denmark and southern Norway, six days further than the auction.

Project website · Documentation · Releases

The day-ahead auction publishes prices for today and, from about 13:00, for tomorrow. Anyone who plans a home battery, an electric car or a heat pump further ahead has to guess. Spot Forecast is a Home Assistant integration that fills the six days after the last published price with an hourly forecast for the DE-LU, NL, DK1, DK2 or NO2 bidding zone (see Supported zones), and says how far to trust it: an expected price per hour, a peak-risk value that the price exceeded in 12–13 % of the hours of the backtest (the target is one in ten), and a low-price value for likely negative prices.

It runs on open data (Energy-Charts and Open-Meteo), needs no account and no API key, and is light enough for a Raspberry Pi: the models are trained every month on GitHub and only evaluated on your device, with numpy.

How it works

1. Known prices. The last 40 days of the zone's day-ahead prices from Energy-Charts, in one request (DK1 and DK2: also DE-LU's, in a second one).
2. Weather forecast. ECMWF IFS (0.25°) from Open-Meteo for the zone's own points and its neighbours' (DE-LU: 10 points in Germany and 12 in neighbouring countries; NL: 7 in and off the Netherlands and 7 around it; DK1 and DK2: 4–5 in Denmark, 7–8 around it and the 22 DE-LU points; NO2: 4 in southern Norway and 8 around it, and the precipitation of six hydro points over the last 90 days and their temperature over the last 30): wind at 100 m, solar radiation, temperature, cloud cover and precipitation. Weather carries most of the skill: without it the error stays close to the naive baseline's.
3. Three tree models. Gradient-boosted trees predict, per hour, the expected price (median), a peak-risk value (P90) and a low-price value (P10), relative to the recent price level, for every hour of the six days after the latest known day.
4. Calibration. Every forecast is kept. The P90 and P10 bands are shifted by how often they were crossed over the last 56 days, so the peak risk stays near one hour in ten. On install, a backfill supplies that history.
5. Output. One price sensor with known and forecast slots, two warning sensors, and long-term statistics of every forecast against the actual price for the look-back card.

A run repeats after each new weather model run (four times a day) and again once the auction result for tomorrow is published. Before that, the forecast covers tomorrow to today + 6; after it, today + 2 to today + 7. Known prices always come first and the forecast follows without a gap. Details: how the model works.

How good is it?

Out of sample, walk-forward: the models were refit month by month on past data only and forecast days they had never seen, with weather forecasts never fresher than, and up to a day older than, what a live install uses. Mean absolute error (MAE) in ct/kWh for DE-LU, from the shipped model's own report (de_lu-2026-10; NL in Supported zones):

Horizon	MAE Apr 2025 – Sep 2026	Naive MAE, same period	Skill, same period	MAE Jul – Sep 2026	Naive MAE, Jul – Sep 2026	Skill, Jul – Sep 2026	Above peak risk
D+1	1.70	3.39	50%	1.99	4.29	54%	12%
D+2	1.78	3.39	48%	2.14	4.29	50%	12%
D+3	1.89	3.39	44%	2.29	4.29	47%	12%
D+4	2.04	3.39	40%	2.50	4.29	42%	12%
D+5	2.31	3.39	32%	2.92	4.29	32%	12%
D+6	2.43	3.39	28%	3.03	4.29	29%	12%

Naive repeats the price of the same hour a week before. Skill is 1 − MAE ÷ naive MAE, which depends far less on how volatile the market was. The model report has the naive MAE for the full period only; for July to September 2026 it is computed from the bundled backfill and Energy-Charts prices (tools/site_data.py), on the same hours as the report's MAE. Above peak risk (full period) is how often the actual price exceeded the calibrated P90; the target is about 10 %.

Peaks come out too low. The forecast gets the timing of price peaks right, not their height. In hours above 21.2 ct/kWh it is on average 7.5 ct too low one day ahead and 9.7 ct six days ahead. For peaks, use the peak-risk band (P90) and the peak-risk sensor that switches on it, not the expected price.

The error depends on the market. From July 2026 gas and electricity prices were higher and more volatile, and the MAE rose to 1.99 ct (D+1) … 3.03 ct (D+6). The naive forecast got worse too (4.29 ct against 3.39 ct), so the skill held at 54 % … 29 %. A frozen model degrades within months when the market shifts, so new models are trained every month and can reach your install without an integration update (see model updates).

The low-price value catches negative prices: in 79–85 % of the hours that turned out negative (depending on the horizon), P10 was below zero. More in accuracy; the website has the forecast against the actual price, per horizon, as an interactive chart.

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗