Impact / Backtest
Belgian solar: our forecast against Elia (Belgian TSO)
134 delivery days of day-ahead forecasts, issued at the same time as Elia's and scored against the same measured output.
14%
lower imbalance settlement cost than Elia's day-ahead forecast
€5.15 vs €5.96 per MWh
€4.9M
saved over 134 days, for the whole Belgian solar fleet
€30.5M vs €35.4M
6%
lower daylight error than Elia's day-ahead forecast
304 MW vs 322 MW
01
Setup
A replay of what a Belgian solar producer would have seen, day by day, with no information from after the forecast time.
- Period
- 1 June to 12 October 2026. 134 delivery days, 12,820 quarter-hours. The model saw none of these days during training.
- Scope
- Belgium's national solar output. The monitored fleet was 12,246 MWp on average and produced 5.93 TWh in the period.
- Issue time
- 18:00 Brussels time on the day before delivery (D−1). Each forecast uses only the weather forecasts and the measured generation that were available at 18:00.
- Benchmark
- Elia's day-ahead solar forecast, published at the same time. Elia is the Belgian transmission system operator (TSO).
- Truth
- Elia's measured solar generation, per quarter-hour. Both forecasts are divided by the same capacity figure, known at the issue time.
02
Metrics
- Daylight MAE
- Mean absolute error: the average size of the forecast miss, per quarter-hour. We count daylight quarter-hours only, because at night both forecasts are trivially zero. We show it as a percentage of installed capacity: 1% is about 122 MW.
- Skill vs Elia
- 1 − our score / Elia's score. We use one definition for two scores: the daylight MAE and the imbalance cost. A positive skill means our forecast does better than Elia's. A skill of +6% on daylight MAE means our error is 6% smaller. A skill of +14% on imbalance cost means our cost is 14% smaller.
- Imbalance cost
-
A producer sells the forecast volume in the Belgian day-ahead auction. Elia then settles the difference between forecast and actual output at its imbalance price. The imbalance cost is the revenue lost against a perfect forecast:
cost = Σ (measured − forecast) × (day-ahead price − imbalance price) × 0.25 h
A positive cost is money lost. We give it in € per MWh of measured output. Prices come from Energy-Charts (day-ahead) and Elia Open Data (imbalance). Trading fees are not included.
03
Results
Cumulative imbalance cost per delivery day: +14% skill vs Elia
Top: the imbalance cost added up day by day, for a producer that trades on each forecast. Bottom: the skill vs Elia on imbalance cost, to date. Over the full period the cost is €5.15/MWh for our forecast and €5.96/MWh for Elia's, a saving of €0.82/MWh.
Daily daylight MAE, ours and Elia's, and the daily skill vs Elia
Top: the daylight MAE of each forecast, per delivery day. Bottom: the skill vs Elia on daylight MAE, for each day. Green bars show a positive skill, orange bars a negative skill.
Measured generation, our forecast and Elia's over time: Belgium
Each day is a separate forecast, issued at 18:00 the evening before. The shaded green ranges show our forecast uncertainty: the measured output falls inside the 90% range 89.8% of the time. Move the slider to step through the period.
Daylight MAE and skill vs Elia per hour after the issue
The error of both forecasts for each hour of the delivery day, counted from the 18:00 issue time: +14h is 08:00. Our skill vs Elia is highest in the morning and evening ramps, when solar output changes fastest.