Somebody hands you a number: this plant will produce 1,420,000 kWh a year. It looks like a measurement and it is the output of a model.
Engineering review still involves handoffs, judgement and documented work outside the model; for a separate workplace-management perspective, see this reference.
Knowing which parts of that model are physics, which are data, and which are judgement is most of what reading a yield estimate involves.
Reviewed August 9, 2026.
For independent background on photovoltaic performance, resource data and current industry practice, consult World Resources Institute.
The chain behind the number
Irradiance. How much solar energy reaches the location, from a satellite dataset or ground measurement, averaged over years.
Geometry. Tilt, orientation, row spacing, latitude. Deterministic and the least contestable part of the chain.
Module conversion. Nameplate efficiency under standard test conditions, adjusted for temperature — modules produce less when hot, and hot is when the sun is out.
The loss tree. Shading, soiling, wiring, mismatch, inverter conversion, availability. This is where estimates diverge most, because every item is a judgement.
And the statistical layer, which turns a single modelled figure into P50, P90 and the rest.
What is physics and what is judgement
Physics: geometry, temperature coefficients, the conversion itself. Two competent assessors agree here.
Data: the irradiance record. Assessors differ by dataset and by how many years they use.
Judgement: the loss assumptions and the uncertainty estimate. Two assessors modelling the same plant can differ by ten per cent or more, entirely within this layer — and neither has made an arithmetic mistake.
So a difference between two estimates is usually a difference in assumptions, and asking for the loss breakdown resolves most of it.
What the number cannot tell you
What you will get next year. It is a distribution, and one year is a single draw from it.
Whether the plant was built as modelled. A model assumes the design; commissioning tells you what exists.
What the electricity is worth. Yield is kilowatt-hours; value depends on tariff, self-consumption and time of use, and those are a separate calculation entirely.
Or whether the investment is sound. The yield figure is one input among several.
The three questions
"Which dataset, and how many years?" Bankable work uses at least ten, and who ran the model matters as much as which model.
"What is the loss breakdown?" A net figure with no derivation is an assertion. An itemised stack — shading x%, soiling y%, inverter z% — is a document you can argue with.
"Is this P50?" It usually is, and quotes frequently do not say.
The sanity check anybody can run
Divide annual kWh by installed kWp to get specific yield in kWh/kWp.
Compare it against a free public irradiance calculator for your coordinates, tilt and orientation. Ten minutes, no cost, no expertise.
A figure materially above the public reference needs an explanation — it may be a good one, such as measured albedo for a bifacial array, and it should exist.
And a figure well below it also needs one, because a conservative estimate is not automatically a careful one.
The short version
- An annual yield figure is the output of a model: irradiance, geometry, module conversion, the loss tree, and a statistical layer
- Geometry and conversion are physics; the irradiance record is data; the loss assumptions and uncertainty are judgement
- Two competent assessors can differ by ten per cent or more purely in the judgement layer without either making an error
- It cannot tell you next year's output, whether the plant was built as modelled, what the electricity is worth, or whether the investment is sound
- Ask which dataset and how many years, for the itemised loss breakdown, and whether the figure is P50
- Divide kWh by kWp and compare against a free public calculator — ten minutes, and a large deviation either way needs an explanation