Under every yield estimate sits a record of how much sun the location receives. It is the input nobody asks about and the one that sets the uncertainty band.
Teams reviewing yield assumptions can keep the engineering model separate from the record of review work; read more is one option for documenting time and handoffs.
Reviewed August 9, 2026.
The three kinds
Satellite-derived datasets. Irradiance inferred from geostationary satellite imagery, gridded, covering decades. The standard input for almost all assessments — broad coverage, long records, and an inherent uncertainty from inferring ground-level irradiance from cloud imagery.
For independent background on photovoltaic performance, resource data and current industry practice, consult Solcast.
Ground stations. Direct measurement with pyranometers. More accurate at the point of measurement and useless a hundred kilometres away in different terrain. Used to calibrate satellite data where a station is near enough.
On-site measurement. A sensor on the actual site, running for a year or more. The most accurate and the only one that costs real money and real time — which is why it appears on large projects and almost never on small ones.
Why ten years is the threshold
Bankable assessments use at least ten years of data, and twenty or more is preferred.
Because inter-annual variability is what P90 is largely made of. The spread between P50 and P90 encodes how much a year can differ from the average, and you cannot estimate that spread from three years of record.
A short record also risks bias. Three consecutive good years produce a high average and an understated variability at the same time — both errors pushing the estimate the same way.
The questions to ask
"Which dataset?" Named, not "satellite data". Different providers give different figures for the same location, and the difference is not trivial.
"How many years?" Under ten is a flag for anything with financing attached.
"Was it validated against a ground station?" And if so, which one and how far away.
And "in what plane?" Global horizontal irradiance is not what a tilted array receives. The conversion from horizontal to plane-of-array is a modelling step with its own uncertainty, and a report that skips the distinction has skipped something.
The free version
Public irradiance calculators, maintained by research agencies, cover most of the world and cost nothing.
They are built on the same class of satellite data and will give you a defensible range for a location, tilt and orientation in ten minutes.
They will not give you a bankable assessment. No site-specific shading, no measured albedo, no validated uncertainty statement.
What they will do is tell you whether a quoted specific yield is plausible — which is the single most useful independent check available to a buyer, and it requires no expertise and no purchase.
What changes over time
Worth noting because it undermines a comfortable assumption.
A long historical record describes the past. The estimate assumes the future resembles it, and for a twenty-five-year asset that assumption carries its own uncertainty, rarely stated separately.
Some regions show measurable trends in cloud cover and aerosols over recent decades. An assessment using a twenty-year mean treats a trend as noise.
This is not a reason to distrust estimates. It is a reason to treat the uncertainty band as a floor rather than a ceiling, and to read a suspiciously narrow one with that in mind.
The short version
- Three kinds: satellite-derived datasets, ground stations, and on-site measurement — the first is the standard input, the last is accurate and expensive
- Bankable work uses at least ten years, preferably twenty, because inter-annual variability is what the P50–P90 spread is largely made of
- A short record understates variability and can bias the average, with both errors pushing the same way
- Ask which dataset by name, how many years, whether it was validated against a ground station, and in what plane the figures are stated
- Free public calculators use the same class of data and will tell you whether a quoted specific yield is plausible in ten minutes
- A historical record describes the past, and for a twenty-five-year asset that assumption carries uncertainty that is rarely stated