Solar · methodology

How this model works, and how wrong it is

Every solar calculator is wrong by some amount. Most do not tell you by how much. This page gives the measured deviation for each kind of site, including the ones where this model performs worst.

For an ordinary residential roof, this model runs +4.2% against PVGIS, with a median absolute error of 5.7%.

Bias is stated first and separately because it matters more than the size of the error. A model that is wrong in a consistent direction makes every payback, saving and return figure downstream wrong in the same direction, and an absolute-error figure alone cannot reveal that.

What it is compared against

PVGIS, the European Commission's Joint Research Centre photovoltaic tool. It is an independent implementation of the same physics, built by a different team from different source data, and it is the reference used across the industry.

144 comparisons run against it: eight sites across five continents — maritime, tropical, arid, high altitude, temperate, arctic, southern hemisphere and equatorial — each at nine orientations from flat to vertical, across two tracks of 72 each. The first track tests the physics against reference climate inputs; the second tests the whole pipeline, our own climate data included. The tables below report the first.

The measured deviation

Positive means this model predicts more generation than PVGIS. Negative means less.

Annual generation, deviation from PVGIS
Class of site	Cases	Bias	Median error	95th percentile	Worst
High altitude, above 1,500 m Thinner air means more direct sun. We currently overpredict here, and the reason is known rather than mysterious: see the open gaps below.	16	+7.9%	9.0%	10.1%	10.1%
High latitude, above 55 degrees Long summer days and very short winter ones. Monthly figures here are the least reliable the model produces, because a small absolute error in a dark month is a large percentage.	14	−1.6%	4.6%	9.3%	9.3%
Ordinary residential roofs The case almost every visitor is in. This is the number to judge the tool by.	32	+4.2%	5.7%	8.9%	9.1%
Vertical and near-vertical arrays Walls and balcony rails. Included deliberately because they are the hardest case and the one most tools quietly refuse. Percentage deviation in near-zero months is dominated by noise.	10	+5.9%	7.2%	52.0%	52.0%

Monthly generation, deviation from PVGIS
Class of site	Cases	Bias	Median error	95th percentile	Worst
High altitude, above 1,500 m Thinner air means more direct sun. We currently overpredict here, and the reason is known rather than mysterious: see the open gaps below.	192	+7.6%	9.0%	14.6%	21.6%
High latitude, above 55 degrees Long summer days and very short winter ones. Monthly figures here are the least reliable the model produces, because a small absolute error in a dark month is a large percentage.	139	−8.8%	7.9%	68.7%	84.1%
Ordinary residential roofs The case almost every visitor is in. This is the number to judge the tool by.	384	+3.0%	6.5%	21.2%	50.5%
Vertical and near-vertical arrays Walls and balcony rails. Included deliberately because they are the hardest case and the one most tools quietly refuse. Percentage deviation in near-zero months is dominated by noise.	114	+12.6%	17.0%	154.4%	289.0%

Monthly figures are considerably worse than annual ones, and that is expected rather than concealed: errors that cancel across a year do not cancel within a month, and a percentage error computed against a dark winter month is large even when the absolute error in kilowatt-hours is small. Use the annual figure to judge a system. Treat a single month as indicative.

What agreement does and does not prove

Agreement with PVGIS means this implementation reproduces an independent implementation of the same physics. That is real evidence, and it is the strongest evidence available without instrumenting roofs.

It is not proof that either model matches your roof. Both share assumptions that could be wrong together: the same clear-sky treatment, the same idea of a typical year, and no knowledge of your particular shading, soiling or wiring. Two models agreeing is weaker evidence than either agreeing with a measured installation, and this page does not claim otherwise.

What is not modelled, and what we know is wrong

These are open gaps. They are listed because a specific, admitted limitation is worth more to you than a general disclaimer, and because each one is a model to be implemented, never a coefficient to be tuned until the numbers look better.

IAM_B0 = 0.05 sits at the low end of the published range for glazed c-Si modules.
PVGIS mountingplace=building models reduced convective cooling that NOCT = 45 may not represent.
The Meinel clear-sky model and the 0.10 clear-sky diffuse coefficient are the weakest constants in the chain and are flagged as such in the specification.
No horizon shading: PVGIS H(h) already embeds terrain horizon effects from its elevation model, ours does not.
Vertical arrays near the equator produce near-zero output in some months, so percentage deviation there is dominated by noise rather than model quality.

The chain, step by step

Sun position — declination, equation of time and hour angle give the sun's altitude and bearing for every interval.
Air mass — how much atmosphere the light crosses, corrected for your elevation.
Clear-sky irradiance — what would arrive with no cloud at all.
Cloud — long-run monthly cloudiness scales the clear-sky figure down to a realistic one.
Beam and diffuse split — direct sunlight and light scattered from the whole sky behave differently and are separated.
Transposition — both components are projected onto your tilted, rotated panel.
Reflection — light striking glass at a shallow angle bounces off instead of entering.
Cell temperature — panels get hot, and hot panels are less efficient.
The loss chain — soiling, DC wiring, mismatch, inverter conversion and downtime, each applied separately so each can be named.

The published figures use 8,760 hourly intervals, every hour of a year. The live preview that responds while you drag a control uses 288 representative intervals, which is why it answers instantly. Where the two would disagree meaningfully, the preview declines to answer rather than answering wrongly.