Solar production forecasting

Forecast accuracy, kept in public.

The headline figure we presented at EU PVSEC 2026, and every forecast issued since go-live, scored against what the plant produced.

How the error is measurednMAE = Σ |forecast − actual| ÷ (hours × AC capacity)

5.1% day-ahead
forecast error.

Model v2 on our reference plant, presented at EU PVSEC 2026: out-of-sample, daytime hours, error as a share of installed capacity. The method, the exclusions and our worst day are published.

DYNVOLT model v25.1%
Our previous model8.3%
Best published research result8.6%
Best of 9 vendors, open utility trial>11%
Read the methodology →See a pilot on your plant
Since go-live, every issued forecast is scored6 Sept – 15 Oct 2026, one bar per day

Cloudy days cost the most. The worst was 20 Sept at 21.4%. Actual output landed inside the forecast's P10–P90 band 80% of the time, which is what an 80% band should do.

How the record is kept.

  • The forecast for each day is issued at 08:45 local, day before delivery (before the 10:00 MEMO gate). Only that first issue is scored; later refreshes never count.
  • Hourly, daytime only (sun more than 5° above the horizon), error as a share of installed AC capacity.
  • Hours with telemetry outages or deliberate curtailment are excluded by fixed rules, listed on the methodology page.
  • The 5.1% headline is model v2 evaluated out-of-sample on our reference plant, as presented at EU PVSEC 2026. The daily bars above are the live issued record since go-live on 6 September 2026.
  • The record so far covers one plant through autumn. Winter will move it, and it stays published either way.

Price forecasts.

Day-ahead price forecasts for D+1 and D+2, with P10 and P90 bands, are issued every morning and frozen at each market's gate for MEMO (North Macedonia), HUPX (Hungary), SEEPEX (Serbia), OPCOM (Romania), IBEX (Bulgaria), HEnEx (Greece) and SDAC Germany. Customers get the full scored history for the markets they trade.

Markets and the trader desk →

Ask any vendor for this page.

A backtest shows what a model could have done. Ask for the record of forecasts it actually issued, day by day.

Request a demoThe forecasting product