Brazil
Solar forecasting
Brazil passed 68.7 GW of installed solar in 2026, a quarter of national capacity, and about 22 GW of it feeds the transmission grid. Between 2022 and 2024 the share of producible renewable energy that never reached the system rose from 0.5 to 9.3 percent. The decision now sits inside the hour, not inside the day.
Solar forecasting in Brazil
Meet us at Intersolar South America. 25 to 27 August 2026, Expo Center Norte, Sao Paulo, booth R3.62. Book a meeting.
Brazil crossed 68,675 MW of installed solar capacity in 2026, 25.9 percent of national generating capacity, according to ANEEL and ABSOLAR. About 22 GW of that is utility-scale plant connected to the transmission grid, and that is where dispatch decisions are taken.
The fleet was built where the irradiance is, in the Northeast and in Minas Gerais, while consumption sits in the Southeast. The lines between the two have not kept pace.
Curtailment first, the settlement price second
Brazil has no penalty on forecast error. A plant does not sell what it produces, it sells the volume it committed to deliver, and the difference settles at the PLD, the short-term price, hourly since January 2021 and specific to each of the four submarkets.
When the operator curtails a plant in the Northeast at midday, the contracted volume is still owed, and it is bought back at whatever that hour is worth. Accountability runs through the price, not through a fine. That is the first thing to understand about this market.
The order of magnitude. A 100 MW plant with a 25 percent capacity factor produces 219 000 MWh a year. A mean forecast error of 5 percent puts 10 950 MWh through the settlement mechanism, which is 2.2 million BRL a year at 200 BRL per MWh, and 4.4 million at 400. The hypotheses are explicit, and the exposure is a function of forecast quality and of price volatility.
Solar in Brazil: the market in figures
Installed solar multiplied by 4.5 in four years, from 14.3 GW in 2021 to 68.7 GW in March 2026, and distributed generation carries about two thirds of it. What changed in 2026 is not the growth, it is the constraint.
The loss is not theoretical. Between 2022 and 2024 the share of producible renewable energy that never reached the system rose from 0.5 to 9.3 percent, concentrated in Bahia, Rio Grande do Norte, Ceara and Piaui.
Three of the five largest solar plants in the country, Janauba, Sol do Cerrado and Helio Valgas, stand in Minas Gerais, where congestion is low. Sao Goncalo in Piaui and Futura 1 in Bahia stand where it is highest. The geography of the risk does not follow the geography of the megawatts.
Indicator
Value
Source
Installed solar capacity
68,675 MW in 2026, 25.9 percent of the mix
ANEEL and ABSOLAR, April 2026
Utility-scale share
About 22 GW, close to one third of solar capacity
ANEEL and ABSOLAR, April 2026
Growth since 2021
From 14.3 GW to 68.7 GW, multiplied by 4.5
ABSOLAR, March 2026
Producible energy never injected
From 0.5 percent in 2022 to 9.3 percent in 2024
ONS, generation restrictions
Settlement price
Hourly since January 2021, four submarkets
CCEE
Free market share of consumption
About 43 percent in 2025
Ministry of Mines and Energy
Battery storage pipeline
Around 18 GW registered for the 2026 capacity reserve auction
Brazilian energy storage association, 2026
Key context
Two forces now act on the same half hour. Curtailment decides how much can be injected, the settlement price decides what the difference is worth, and battery storage turns both into a decision. All three resolve inside the hour, which is exactly where a satellite-based forecast has the least to say.
Why a tropical sky defeats satellite-only forecasting
Convective cells, not fronts
Most of Brazil sits in an A type climate. Clouds form by surface heating over a few square kilometres, develop vertically and dissipate in place. There is little coherent horizontal transport, and horizontal transport is exactly what cloud motion methods extrapolate.
At mid latitudes a front is a large organised structure a satellite can track for about three hours. A tropical convective cell is born, peaks and is gone inside the hour. The gap is measurable: on a Brazilian site a GOES estimate read 1,049 W/m2 in the same ten minutes during which the on-site pyranometer measured a minimum of 229 W/m2. The satellite was not wrong about the region, it was blind to the cell over the plant.
Obligation
What it covers
Why it matters to a forecast
Obligation
What it covers
Why it matters to a forecast
Programa de geracao, D-1
Expected output for the next day at 30-minute resolution, 48 values
The grid you produce on, and the grid your error is read on
Intraday updates
Revision whenever weather or availability moves
The value sits in the revision, not in the first submission
Real-time telemetry
Active power, voltage, frequency, plant state, through SINtegre
The same instruments serve monitoring and forecasting
Key context
There is no general obligation yet to submit a standardised forecast curve. The operator builds its own view with its own meteorological models, revised in 2025 in submodule 2.4 of the Procedimentos de Rede, and requirements are tightening on real-time measurement. Planning ahead of that curve costs less than following it.
Where the fleet is, and under which sky
The large plants sit in the Northeast and in Minas Gerais, and almost all of the country lies in an A type climate, equatorial or monsoon, with the semi-arid sertao in between. These are convective cloud regimes, not the frontal systems most forecasting models were built for.
Region
Dominant constraint
Forecasting difficulty
Region
Dominant constraint
Forecasting difficulty
Bahia and Piaui
Very high curtailment, transmission congestion
Semi-arid, but the same convective cloud structures
Rio Grande do Norte and Ceara
Wind and solar competing for the same lines
Coastal convection, cells forming over a few kilometres
Minas Gerais and Sao Paulo
Low congestion, the largest utility-scale plants
Summer monsoon regime, high day-to-day variability
Key context
Brazilian cloud physics is not mid latitude cloud physics. A method validated in Europe on advected fronts carries an implicit assumption that does not hold over the sertao, and the error appears precisely in the horizons where curtailment and storage decisions are taken.
What we measured on a Brazilian site
Ninety-six days, from 1 May to 4 August 2026, on a Brazilian utility-scale plant. A forecast issued every fifteen minutes, compared with the output the plant actually delivered.
At the two-hour horizon, the forecast built on on-site data reduced mean absolute error by 25 percent against a numerical weather model, and by 24 percent against persistence. Below 21 minutes persistence from a local sensor still wins, and only a sky imager network adds information. Adding sky imagers to a site already equipped with pyranometers improved skill by 20 percent below thirty minutes.
Horizon
What the market uses today
Where the gap sits
Horizon
What the market uses today
Where the gap sits
Day ahead, D-1
Numerical weather prediction
Adequate at plant level, this is the programme the operator consolidates
Intraday, one to four hours
Weather models and satellite imagery
Local convection not resolved, revisions arrive late
One minute to one hour
No effective coverage
Cells live and die below satellite resolution
Pre-construction resource
Satellite databases only
Tropical bias not calibrated, and the P90 carries it
Key context
The 21-minute threshold is the useful number. It says where instruments stop being a refinement and become the only source of information. Any storage strategy arbitrating the hourly PLD, and any attempt to anticipate a curtailment order, lives below that threshold.
CalibSun in Brazil: on-site data where the models stop
Satellite models and numerical weather prediction describe the region. They do not describe the plant. CalibSun instruments the site, calibrates against what the plant actually receives, and forecasts on the grid the market settles on.
INSTANT, nowcasting below the hour, where curtailment and storage decisions are taken
INSTANT deploys hemispherical sky imagers and pyranometers on the plant. Cloud identification, cloud height, cloud motion and the projection of the shadow on the plant layout are computed on site, continuously. Output forecast from one minute ahead, at the timescale a convective cell actually lives, with 20 percent additional skill below thirty minutes measured on a site already equipped with pyranometers.
NEXT, intraday and day ahead on the thirty-minute grid the operator uses
NEXT assimilates plant measurements, satellite imagery and weather models, and returns the half-hourly curve expected for the next day, plus intraday revisions when the sky moves. On a Brazilian site over 96 days, mean absolute error at two hours fell by 25 percent against a weather model alone.
FUTURE, one year of on-site measurement before the plant exists
FUTURE runs twelve months of on-site measurement at five-minute resolution to calibrate the long-term satellite record, which carries a bias in tropical climates. A P50 and a P90 built on a calibrated record are the ones a lender can price, and the same instrumentation later separates curtailment from soiling in operation.
Need more information?
Get in touch!
- Questions
Frequently Asked Questions
Is a solar production forecast mandatory in Brazil?
Plants connected to the transmission grid and supervised by the operator submit a production programme for the next day at a thirty-minute resolution, together with availability, planned outages and real-time telemetry, through the SINtegre portal, with intraday updates when conditions move. There is no general obligation to deliver a standardised forecast curve, and the operator builds its own estimate with its own meteorological models. Requirements are tightening, on real-time measurement in particular.
How is a forecast error paid for in Brazil?
Not as a penalty. Brazil has no national mechanism penalising forecast error. Deviations between the contracted volume and the injected volume settle at the PLD, the short-term settlement price, hourly since January 2021 and specific to each of the four submarkets. The cost is a market exposure, not a fine, and it grows with the volatility of that price.
Why do satellite-only forecasts underperform in Brazil?
Because tropical convection does not advect. Cells form over a few square kilometres, peak and dissipate inside the hour, so extrapolating cloud motion from imagery has little to extrapolate. On a Brazilian site a GOES estimate read 1,049 W/m2 while the on-site pyranometer measured a minimum of 229 W/m2 in the same ten minutes.
Below which horizon does on-site measurement become necessary?
Twenty-one minutes, measured over 96 days on a Brazilian utility-scale site. Below that horizon persistence from a local sensor beats the models, and only a sky imager network adds information. Adding sky imagers to a site already equipped with pyranometers improved forecast skill by 20 percent below thirty minutes.
Regulatory and market information. The regulatory information on this page describes the Brazilian framework as of August 2026. Grid operation rules are set by the ONS in the Procedimentos de Rede, notably submodule 2.4 revised in 2025 and module 4 on operation programming. Market rules and the settlement of differences are set by the CCEE, and tariff and connection rules by ANEEL. Requirements change. Please verify the applicable provisions with the ONS, the CCEE and your distribution concessionaire before taking any commercial or operational decision. CalibSun accepts no liability for any use made of this information.
Sources. Installed capacity and the split between distributed and utility-scale generation: ANEEL and ABSOLAR, April 2026. Restrictions on renewable generation between 2022 and 2024: ONS. Hourly settlement price and the settlement of differences: CCEE. Share of consumption on the free market in 2025: Ministry of Mines and Energy. Battery storage pipeline for the 2026 capacity reserve auction: Brazilian energy storage association. Largest operating solar plants: ANEEL, August 2026.
CalibSun measurements. Forecast performance figures come from an evaluation over 96 days, from 1 May to 4 August 2026, on a Brazilian utility-scale plant, with a forecast issued every fifteen minutes and compared with delivered output. Skill scores are stated against the reference named in each case, a numerical weather model or persistence, at the horizon named. Satellite comparison: GOES estimates at ten minutes against an on-site pyranometer at five minutes, 2 and 3 August 2026.
