The model reads the ONS curtailment history and projects the next 72 hours, half-hour by half-hour. No number lands here without its own error beside it — the yardstick and the method live in the methodology, and the hit rate is measured in public on the accuracy page.
At this page's build, the first day's typical error was 2,537 MW — measured over the last four weeks, the same way operations forecast.
Real data throughSeptember 26, 2026 at 11:30 PM
Everything past this instant is model — and the page says how much it errs.
Latest forecastSeptember 28, 2026 at 7:00 AM
The model runs twice a day — -3h ago, at this page's build.
The three days, as a curve
The model's forecast (the line) and the band where 80% of realized values tend to fall (the shaded area) — day 1 rests only on real data, days 2 and 3 are recursive. Hover the line to see each half-hour's value.
Model forecastP10–P90 band (80%)
The next three days
Sunday, September 27
Mean forecast curtailment
4,387 MW
Forecast peak
9,669 MW at 13:30
Energy over the day
105,281 MWh
Horizon's typical error
±2,537 MW
Uncertainty band (P10–P90)
0–9,596 MW
Monday, September 28
Mean forecast curtailment
5,944 MW
Forecast peak
12,868 MW at 12:30
Energy over the day
142,663 MWh
Horizon's typical error
±3,001 MW
Uncertainty band (P10–P90)
0–15,805 MW
Tuesday, September 29
Mean forecast curtailment
6,959 MW
Forecast peak
16,918 MW at 09:30
Energy over the day
167,012 MWh
Horizon's typical error
±3,111 MW
Uncertainty band (P10–P90)
0–21,538 MW
The first day is forecast with real data only. The second and third are recursive — the model rests its lags on what it said the day before — which is why the typical error grows with the horizon. Uncertainty is not a footnote: it travels inside the file itself. The P10–P90 band is the range where 80% of the realized value tends to fall — two extra models, tuned to forecast the quantile instead of the mean, with a correction that guarantees that coverage (the math is on the accuracy page).
Field by field
The forecast is public at the same address this page reads: nameplate.com.br/previsao.json. The field contract is locked by tests on both sides of the bridge:
snapshot_at
When this forecast was generated.
dado_ate
The newest instant of real data used — everything after it is model.
modelo
The variant in production. A new variant gets a new name — and a methodology note.
fator_nivel
The scale recalibration applied, measured on the last four weeks' error.
mae_30d_mw
Each horizon's typical error (1, 2 and 3 days), measured the way operations forecast.
mae_30d_regua_mw
The same error, for the yardstick (repeating the last observed day), on the same instants — the model's edge, checkable.
vies_30d_mw
Each horizon's signed error: positive is overestimating, negative is underestimating.
cobertura_80_pct
The fraction of realized values that fell inside the published P10–P90 band, measured on the same trials. Nominal is 80%.
margem_p80_mw
The correction that widens the two quantile models' raw band until coverage hits 80% — the cost of promising it right.
previsao
The 144 half-hours of the next three days: previsto_mw and each one's p10_mw–p90_mw band.
These numbers are baked when the site is generated; the always-current value lives in previsao.json itself. The curtailment math and the model's premises are in the methodology.