Solar irradiance databases for long-term evaluation: SARAH-3, CAMS and NASA POWER

In photovoltaic solar energy projects, the estimate of GHI (Global Horizontal Irradiance) of the long term is the most critical input of energy generation — corresponding to what the average wind speed represents in wind projects. Just as the choice of atmospheric reanalysis affects wind MCP (NT.02), the choice of the irradiance database directly affects the estimated P50 generation and, by cascade, the entire uncertainty structure that determines P90 for DSCR purposes.

Three free databases dominate the Brazilian market: SARAH-3 (CM SAF/EUMETSAT/DWD), CAMS (Copernicus Atmosphere Monitoring Service) and NASA POWER (derived from the MERRA-2 reanalysis). Each one has radically different architecture, resolution and performance — and systematic errors different by region of Brazil. This note presents a methodological and quantitative comparison to guide the selection and use of these sources in bankable reports.

The three databases: architecture and resolution

The differences in performance between the bases are a direct consequence of their derivation architectures. Knowing the principles of each one is fundamental to understand where each base fails and why:

CM SAF SARAH-3 (Surface Solar Radiation Data Set — Heliosat, version 3). Image derived from geostationary METEOSAT satellites (EUMETSAT) by the European Space Agency and DWD. Covers Europe, Africa and South America from 1983 to 2022. Spatial resolution ~0.05° (~5.5 km) and temporal time. The Heliosat method converts reflection of the top of the atmosphere into cloud coverage index (Cloud Index, CI) and then into GHI via relation to clearsky radiation. The main weakness is the sensitivity to aerosol: in regions with high aerosol load (such as the Northeast semi-arid, affected by transocean Saharan dust), the Heliosat method tends to interpret part of the attenuation by aerosol as clearsky, resulting in overestimation of GHI.

CAMS (Copernicus Atmosphere Monitoring Service). Developed by ECMWF and produced by MINES ParisTech. Combines high quality clearsky model (Ineichen-Perez calibrated) with aerosol and cloud data from CAMS itself. Space resolution similar to SARAH-3 (~2–5 km via METEOSAT satellite), with hourly data available since 2004. Because of its explicit and calibrated treatment of aerosol and water vapour, it tends to perform better than SARAH-3 in regions with high aerosol load. Its shorter historical period (from 2004) limits the estimate of long-term interannual variability.

NASA POWER. Derived from the atmospheric reanalysis MERRA-2 (50 km of spatial resolution) with satellite corrections. Available since 1981, global coverage, totally free of charge and with robust API. It is the most used for convenience — but has the worst performance among the three bases for local applications, with RMSD typically 20-32% and systematic positive bias in Brazil (superestimate GHI), direct consequence of the smoothing of coarse spatial resolution that does not capture variability of local cloud cover.

Regional Bias in Brazil: MBD by database

The Mean Bias Difference (MBD) — defined as (satellite − piranometer) / piranometer — systematically varies by region according to cloud cover, aerosol and local topography. Validation studies with SONDA (INPE) and BSRN stations reveal three distinct patterns:

Northeastern Brazil (semiarid): High insolation, low cloud cover, but significant load of Saharan dust aerosol transported by the Atlantic. SARAH-3 overestimate (+3–5%) because the Heliosat does not separate well attenuation by clearsky aerosol. CAMS performs better (+0.5–2.5%) by the explicit treatment of aerosol. NASA POWER presents the highest positive bias (+3–7%) and high RMSD.

South-East and Central-West (closed and transition): Mixed regime — dry in winter, with intense convection in summer. Cu/Cb fast development cloud cover makes it difficult for all models based on geostationary satellite. SARAH-3 is close to zero bias in winter and slightly positive in summer (+0.5–2%). Similar CAMS. NASA POWER maintains positive bias (+1–3%) throughout the year.

South (subtropical, frequent cold fronts): Increased cloud cover variability, jet winds, extra-tropical fronts. SARAH-3 tends to underestimate slightly (−2 to 0%) because the calibrated IC algorithm for tropical convective clouds adapts worse to frontal systems. CAMS is more robust. NASA POWER keeps bias slightly positive (+0.5–2%).

Seasonality: how bias vary throughout the year

The average annual error hides important seasonality. In solar energy projects, the monthly profile of GHI determines the seasonal dispatch and is critical for energy analysis at peak times (HFP vs HCP). Databases with seasonal systematic error — even if the annual MBD is small — can distort the estimate of revenue in PPP structures with differentiated price per hour of the day or season.

For the Brazilian Northeast, there is a common pattern: NASA POWER overestimates more in the dry period (June–October), when Saharan dust reaches peak concentration and the MERRA-2 algorithm does not adequately resolve aerosol attenuation. SARAH-3 shows seasonality linked to the convective cloud cycle — closer to the soil in the wet period (January–April). CAMS, by explicitly treating aerosol, has more stable seasonal behavior.

Long-term GHI uncertainty and impact on P90 FV

The uncertainty of long-term GHI for bankable report purposes combines two terms:

• Satellite model uncertainty (σ_sat)The MBD and RMSD of the database versus the physical reality. It can be reduced (but not eliminated) by validation with soil piranometer. If the database overestimates 3% and the analyst does not correct, this bias enters directly into the P50.
• Interannual variability of GHI (σ_IAV)The natural fluctuation of insolation year by year, caused by ENSO, Atlantic dipole and other climate variability modes. For Brazil, σ_IAV is typically 2–4% of the average GHI for periods of 20–30 years. It cannot be eliminated — only estimated.

σ_total for the estimation of long-term GHI, propagated via RSS, determines the FV generation P90 (together with uncertainty of conversion model, losses, degradation). A database with high σ_sat — such as NASA POWER — contributes to almost double the uncertainty that SARAH-3 or CAMS, even for an analyst who corrects the average bias, because the residual variability (monthly RMSD) is still higher.

Selection hierarchy for bankable reports

Based on performance analysis and typical values of σ, the source selection hierarchy for long-term GHI estimation in solar resource reports is:

1. Calibrated soil pyronometer + satellite base for long-term extrapolation. Required for projects above 50 MWp in due diligence. The piranometer corrects the bias of the satellite base and reduces σ_sat to ~1.0–1.5%; the satellite base provides the long time series to estimate interannual variability. NT.03 For QA/QC criteria of the ground station.
2. The SARAH-3 + CAMS set. Suitable for preliminary due diligence, feasibility studies and minor projects. Combining the two bases is not simple arithmetic mean — the correct process is: verify the MBD of each base against the normal regional climatological (BSRN or SONDA), individually correct the biases, then calculate the weighted average by the reverse variance. Typical result: σ_total ~2.5%.
3. CAMS (alone, with regional bias correction). – Acceptable if there is documented evidence of low MBD for the specific region and the historical series is sufficient (≥ 15 years). It should be declared as a single source with its corresponding uncertainty.
4. SARAH-3 (only, NE excluded). Suitable for SE/CO and South. For NE, aerosol bias should be corrected with regional AOD (Aerosol Optical Depth) data from AERONET or CAMS, or the base should be replaced by CAMS.
5. NASA POWER — not recommended as a primary source. – Acceptable only for initial screening (pre-viability analysis) or as input data for re-analysis models. It should be explicitly declared as a limited source with σ ≥ 5.5%. Reports containing P50 derived exclusively from NASA POWER without bias correction are systematically returned by qualified ITAs.

Overlay period and validation window

The validation of a satellite base against soil piranometer requires a sufficient overlap window to capture seasonality and variability. Minimum criteria:

• 12 continuous months At a minimum absolute — captures a full seasonal cycle but not annual variability;
• 24 months — suitable for calculating monthly MBD with an acceptable confidence interval (±0.5–1.0% for R² ≥ 0.95 on a time scale);
• 36 months or more — sufficient to estimate local σ_IAV and calculate the long-term correction factor with uncertainty below 1.5%.

Overlay periods of less than 12 months should not be used to correct the satellite base bias in due diligence reports. The validation window should be documented in the report with: start and end dates, availability of pirometer data (% of hours valid per month, according to IEC 61724-1), and monthly and total MBD/RMSD table.