What 15.3 kW actually means

No roof delivers it.
This one spreads its panels across four faces of a hip roof. That costs 20% of the total before any electricity is generated.
Of the three deductions between the rating and the meter, the smallest is the loss even the best days have.
Over 533 days the array delivered 42% of what those 36 panels would make due south at 37° under a sky that is always clear.

Where the other 58% goes

Start with the contract: 36 panels at 425 W, all due south at 37° under a sky that is never cloudy. My roof does not do that.
Each step down is a different reason. Only the last of them is weather.
The first step is the roof itself. Geometry alone settles it: no measurement, no year of data, only where the four faces point.
The second step is what even the best days lose: wiring, the inverters, and shading, which includes the neighbor's tree over the west face and changes with its leaves.
The third step is the sky, with any panel heat beyond what the best days had.

The model under the chart shows the first two steps. It is a schematic, and the figures beside it are the cascade's.
Its first toggle puts every panel on one plane, due south at the baseline's tilt, with no tree. So the second removes the tree from my four faces only.
It opens at the moment the Sun Position page opens on, when the tree shades the most panels. You can drag it to turn it.

This control zooms the charts that have a date axis. Totals are for the full record.

a day's measured share of its own clear-sky ceiling the best-day envelope

Each day is set against what a clear sky would have given that day, at that sun angle, on these four faces. So the seasons are already divided out.
A good December day and a good June day both land near the top, though the June day made far more energy.
The flat top edge is the useful part. Its line is the average of the five best days, 81%; the best single day reached 80.9%.
Below the edge, the shortfall is mostly cloud.
The points on the floor from 25 January to 3 February 2026 are the 10 days in a row the array made under 1 kWh, with snow on the panels.
Every point is a day from the meter's cumulative counter. The days before the counters began are left out, because no counter can check their best days.
Python computes the clear-sky ceiling for my location rounded to the nearest 0.5°, a point up to about 35 km from the house.
The rounding moves the percentages on this page by a few tenths of a point. That is the price of not publishing where the house is.

This is the same days as a distribution. It has no useful average.
The mass sits at both ends: clear days against the ceiling, and overcast days near the floor.
Little sits in between, so the mean of 50% describes very few days.

Each day is already measured against its own ceiling. So this is not a chart of how much sun there was.
It shows how much of the available sun the array collected.
That is a question about cloud, not about the calendar.

The same 36 panels, modeled under the same clear sky, are arranged three ways.
A hip roof spreads them over four faces. Two of those faces point north-north-west and east-north-east.
Laying every panel flat would collect more than this roof does.
A shallow tilt pointed away from south is worse than no tilt at all.