Solar irradiance and resource data: sources and uncertainty

Every energy yield assessment, financing model, and payback calculation for a PV project ultimately rests on an irradiance estimate for a location where, in almost all cases, nobody has ever installed a calibrated ground-based sensor. Getting this input wrong by even a few percent compounds through capacity factor, revenue, and debt sizing for the life of the asset. This article is for engineers and analysts who specify or interpret resource data for feasibility studies and bankable energy yield assessments, and who need to understand not just where to download a dataset but how much to trust it. We cover the three irradiance components, the satellite and ground datasets in common use, typical year (TMY) construction, and how uncertainty is expressed and propagated into the P50/P90 exceedance figures that lenders require.

The three irradiance components

Global Horizontal Irradiance (GHI) is total solar radiation received on a horizontal surface, combining the direct beam and the diffuse sky component. Direct Normal Irradiance (DNI) is the beam radiation received on a surface always held perpendicular to the sun, relevant for tracking systems and concentrating technologies. Diffuse Horizontal Irradiance (DHI) is the scattered component reaching a horizontal surface from the rest of the sky, excluding the direct beam. The three are related by GHI = DNI x cos(theta_z) + DHI, where theta_z is the solar zenith angle. For fixed-tilt PV systems, plane-of-array (POA) irradiance is derived from GHI, DNI, and DHI via a transposition model (e.g. Perez, Hay-Davies, or isotropic), which estimates how much of the diffuse sky dome and ground-reflected radiation reaches a tilted surface. Transposition model choice itself introduces 1-3 percent uncertainty in POA estimates, larger at high tilt angles and high latitudes where diffuse fraction is greater.

Why the split matters for technology choice

Sites with a high DNI fraction (clear, dry climates such as parts of the Middle East, southern Spain, or northern Chile) favor single-axis tracking, since trackers capture the directional beam component efficiently; the marginal energy gain from tracking is much smaller in climates with high diffuse fraction (northern Europe, tropical monsoon regions), where GHI is dominated by DHI and a tracker adds cost without proportional yield benefit.

Measurement and modeling sources

Ground stations

Calibrated ground stations using thermopile pyranometers (per ISO 9060, spectrally flat secondary-standard instruments) remain the reference against which all modeled data is validated. Well-maintained stations achieve measurement uncertainty of 2-3 percent for GHI on an annual basis. The limitation is coverage: most project sites are tens to hundreds of kilometers from the nearest long-term, well-calibrated station, and local microclimate effects (coastal fog, orographic cloud, urban aerosols) are not captured by a distant station.

Satellite-derived databases

Satellite-derived irradiance, produced by processing geostationary satellite imagery through a radiative transfer or empirical cloud-index model, is now the primary source for bankable resource assessments because it provides spatially continuous, multi-decade time series anywhere within satellite coverage. Commonly used databases include PVGIS (JRC, European Commission, using the SARAH and ERA5 datasets, strong coverage across Europe, Africa, and parts of Asia), NSRDB (NREL, USA, National Solar Radiation Database, covering the Americas at high resolution), CAMS Radiation Service (Copernicus Atmosphere Monitoring Service, global coverage), and SolarGIS and Solargis-derived commercial products used widely in bankable due diligence. Satellite-derived GHI typically carries an uncertainty of 4-8 percent on an annual basis depending on climate (lower in clear, stable climates; higher in regions with frequent broken cloud or high aerosol loading, such as parts of Southeast Asia or dust-affected desert regions).

Typical Meteorological Year (TMY) construction

A TMY is a single synthetic year built by selecting, for each calendar month, the most statistically representative month from a multi-year historical record (typically 10-30 years), based on indices covering irradiance, temperature, wind speed, and humidity. It is not the average year and is not any single year that actually occurred; it is a composite designed to represent long-term typical conditions for simulation purposes, most commonly consumed by yield modeling tools such as PVsyst or SAM. The method used to select representative months matters. The widely used Sandia/NREL method weights GHI most heavily, but different tools weight temperature and wind differently, so two TMY files built from the same underlying multi-year record by different providers can show measurable differences in resulting simulated annual yield, typically within 1-2 percent for GHI-related parameters.

Interannual variability versus TMY smoothing

A TMY intentionally suppresses year-to-year variability that a real project will experience. Actual annual GHI at most European sites varies by roughly 4-8 percent (one standard deviation) from year to year due to cloud cover variability; desert climates show lower interannual variability (2-4 percent) because clear-sky conditions dominate most days. This variability, not the TMY's central estimate, is what drives the P50/P90 exceedance analysis described below.

Uncertainty and exceedance probabilities

A P50 estimate is the annual energy yield expected to be exceeded in 50 percent of years - effectively the median (or mean, depending on methodology) long-term expectation. A P90 estimate is the yield exceeded in 90 percent of years, i.e., a conservative, low-probability-of-shortfall figure used for debt sizing, since lenders want confidence that scheduled debt service will be met in all but the worst 10 percent of years. P90 is derived from P50 by combining the interannual variability of the resource with the combined uncertainty of the yield model, expressed as a standard deviation, then applying a one-sided normal (or sometimes different) distribution: P90 = P50 x (1 - 1.28 x sigma), where 1.28 is the one-sided normal z-score for the 90th percentile and sigma is the combined coefficient of variation.

Worked example

A site has a P50 annual yield estimate of 1,650 kWh/kWp from a bankable energy yield assessment. The assessment quotes an interannual GHI variability of 5.5 percent and a combined modeling uncertainty (irradiance data, transposition, system losses) of 4.0 percent. Combined uncertainty sigma = sqrt(5.5^2 + 4.0^2) = sqrt(30.25 + 16.0) = sqrt(46.25) = 6.8 percent. P90 = 1,650 x (1 - 1.28 x 0.068) = 1,650 x (1 - 0.087) = 1,650 x 0.913 = 1,507 kWh/kWp. A P10-year, if lenders instead needed an upside case, would use +1.28 sigma: 1,650 x 1.087 = 1,794 kWh/kWp. This 287 kWh/kWp spread between P90 and P50 (about 17 percent) is a typical order of magnitude for a temperate European site and directly affects how much debt a project can support.

Comparing common resource datasets

Site adaptation and bias correction

Bankable assessments typically apply "site adaptation": correcting the long-term satellite time series using a shorter period of on-site ground measurement (commonly 6-24 months from a temporary met mast or reference cell), which can reduce effective uncertainty by 1-2 percentage points relative to using satellite data alone. This is standard practice for utility-scale projects above roughly 10 MWp but often skipped for smaller commercial rooftop projects where the cost of a temporary met station is not justified relative to project value.

Common mistakes

• Using a TMY file directly as a single-year production forecast rather than a modeling input, then being surprised when actual output in any given year differs by several percent.
• Applying P50 yield uncertainty as if it were a single number across all technologies, when tracking systems and bifacial modules carry additional model-specific uncertainty beyond the base irradiance uncertainty.
• Mixing GHI-based and POA-based capacity factor figures when comparing sites, without accounting for transposition model differences.
• Ignoring seasonal distribution when only annual GHI is compared - two sites with identical annual GHI can have very different summer/winter production profiles relevant to self-consumption or seasonal tariffs, see also self-consumption calculator.
• Treating satellite-derived irradiance as ground-truth in regions with known high aerosol or complex terrain effects, where uncertainty is materially higher than the global average figures often quoted.
• Failing to reconcile the resource dataset's time zone and averaging convention with the PV simulation tool, causing subtle timestamp misalignment in sub-hourly analyses.

Frequently asked questions

Which dataset should I use for a bankable yield assessment?

Most bankable assessments for utility-scale projects combine a long-term satellite dataset (PVGIS, NSRDB, or a commercial provider) with a period of on-site ground measurement for site adaptation. For early-stage feasibility work, PVGIS or NSRDB alone is generally sufficient and free to access.

How much does POA irradiance differ from GHI at a typical European fixed-tilt site?

At a 30-35 degree optimal tilt in central Europe, POA irradiation is typically 10-15 percent higher than GHI on an annual basis, though this varies with latitude, diffuse fraction, and ground albedo.

Is P90 always more conservative than P50?

Yes, by definition P90 represents a yield level with a 90 percent probability of being exceeded, so it is always equal to or lower than P50 under standard exceedance methodology. Some contracts also specify P75 or one-year P90 versus ten-year P90 (which accounts for cumulative variability across a loan term), so it is important to confirm exactly which exceedance definition a given figure uses.

Do TMY files account for climate change trends?

Traditional TMY construction from historical records does not explicitly project forward climate trends; it represents historical typical conditions. Some newer resource products incorporate adjusted or forward-looking climate data, but these are not yet standard practice across the industry and should be flagged explicitly when used in an assessment.

How does soiling interact with irradiance data in a yield assessment?

Irradiance data describes the resource available at the top of the array; soiling losses are modeled separately as a system loss factor, typically 1-5 percent annually in temperate low-dust climates and up to 10-15 percent in arid, high-dust regions without regular cleaning. Combining an already-conservative P90 irradiance estimate with an overly optimistic soiling assumption is a common source of yield assessment error.

Standards and references

IEC 61724-1: Photovoltaic system performance - Part 1: Monitoring (defines irradiance measurement requirements referenced in resource assessment). ISO 9060: Solar energy - Specification and classification of instruments for measuring hemispherical solar and direct solar radiation. IEC 60904-3: Photovoltaic devices - Part 3: Measurement principles for terrestrial photovoltaic solar devices with reference spectral irradiance data. Reference material: NREL National Solar Radiation Database documentation; JRC PVGIS methodology reports (European Commission); IEA PVPS Task 16 reports on solar resource for high penetration and large scale applications; IRENA renewable resource assessment guidance. See also system sizing calculator for translating a chosen resource estimate into expected annual output, and PV monitoring and performance ratio for how measured irradiance feeds into ongoing performance tracking once a plant is operating. Background on the glossary covers additional resource assessment terminology.