SolarSite Advisory

Site selection for utility-scale solar is well served by public data. Irradiance is published, production is modelled, and a great many dashboards already rank counties by how much sun they receive. I was interested in the question that follows rather than the one that precedes it: once an array is standing, what can still be grown underneath it, and is the answer ever material enough to change where a project should go?

The appeal of that question is that it has a real answer in the literature, and almost no one has connected that literature to county-resolution siting data.

Thirteen of thirty-nine crops tolerate shade. Seven yield more under panels than in open field.

The intuitive model is that shade costs yield, and for most crops it does. But a peer-reviewed systematic review of agrivoltaic field trials classifies 13 of its 39 crops as shade tolerant, and records 7 of them yielding more under partial shade than in open ground. Chicory comes in 69 percent heavier under moderate shade; alfalfa gains 30 percent, common beans 25 percent, lettuce 18 percent.

The mechanism is not exotic. Partial shade lowers canopy temperature and evaporative demand, and for species that are not light-limited in high summer, the relief outweighs the loss of radiation. What makes the finding useful is that it inverts the framing: dual use is not uniformly a sacrifice, and the size of the sacrifice, where there is one, varies enormously by crop.

Every figure in the report resolves to a public, citable source. Nothing about the physical or agronomic model is asserted.

• NASA POWER Monthly climatology for all 3,142 county centroids: global horizontal irradiance, photosynthetically active radiation, temperature, precipitation. 37,704 records.
• NREL PVWatts v8 Modelled AC production per megawatt DC, every county in three array geometries. 9,426 API calls.
• USDA NASS County crop yields, state prices received, and county pastureland cash rent as a grazing revenue proxy.
• Energies 2026, 19(2), 539 Systematic review of experimental agrivoltaic trials, supplying every yield-retention factor with its own confidence grading.

Production was validated against the source it came from. Scaling factors derived for two array geometries land within 3.9 and 4.7 percent of PVWatts’ own modelled output for those geometries. Applying the single-axis tracking factor to the fixed-tilt baseline gives a national capacity factor of 23.0 percent on an AC basis, against a fleet median of 24 percent in Berkeley Lab’s empirical utility-scale survey.

Interpolating on measured light transmission, rather than snapping to three shade bands.

The review reports retention at three shade levels, anchored at 85, 67.5 and 50 percent ground-light transmission. The six array geometries in the model sit at 35, 45, 55, 65, 72 and 85 percent, so only one of them lands on an anchor. The obvious implementation assigns each geometry to its nearest band.

That implementation is optimistic, and it is optimistic in the worst available direction. Snapping overstated retention on 102 of 234 crop-by-geometry pairs, and the error concentrated on precisely the elevated designs the report exists to evaluate. Corn under an elevated fixed-tilt array reads 1.00 when snapped. Interpolated linearly on measured transmission, the same crop and geometry returns 0.714.

A dataset that was complete, plausible, and quietly wrong.

NASA POWER serves irradiance in megajoules per square metre per day to its agricultural community, and in kilowatt-hours to its renewable energy community. The fetcher requested the agricultural community, correctly, because photosynthetically active radiation is published only there. It then labelled the column as kilowatt-hours and applied a plausibility filter admitting values between 0 and 15.

Read as megajoules, that filter silently discarded every month sunnier than 4.17 kWh per square metre per day. It did not fail. It did not warn. It produced a well-formed CSV in which one county out of 3,134 retained all twelve months, and the surviving record was systematically biased toward winter.

The tell was arithmetic rather than intuition: rows per county came to 5.5 where the schema requires 12. Confirming it took one call to the API’s own unit metadata, which reports 17.05 MJ and 4.7357 kWh for the same coordinate, a ratio of exactly 3.6. After conversion the series spans 2.29 to 5.96 kWh per square metre per day, Arizona highest and Alaska lowest, and the PAR-to-irradiance energy ratio settles at 0.455 against a physical constant near 0.45 to 0.50.

I am reporting this because it is the most instructive thing that happened during the build. A pipeline that throws an exception costs an afternoon. A pipeline that returns a complete-looking artifact with half its data quietly removed will survive to the end of a project, and the only defence is checking row counts against what the schema requires before believing anything downstream.

Two of the six array geometries, conventional fixed tilt at 35 percent transmission and single-axis tracking at 45 percent, sit in denser shade than any trial in the review. Their retention values are clamped extrapolations, not measurements, and the report says so on the page whenever such a geometry is selected.

This is worth stating plainly rather than burying, because the two extrapolated geometries are also the two most commonly built. Any crop figure the report gives for a conventional array should be read as a floor derived by extension, not as an observation.

Vertical bifacial admits 2.4 times the ground light of a conventional array, and delivers a third of the energy per acre.

The trade-off is monotonic and steep. Geometries that let light reach the ground do so by spacing panels apart, and spacing panels apart reduces the megawatts an acre will hold. Output per megawatt is the wrong instrument for a land question, and it inverts the ranking: an elevated wide-row tracker produces 16 percent more per megawatt than an elevated fixed tilt, yet fits a third fewer megawatts on the acre, and so delivers 23 percent less energy from it.

Whether that trade is worth making is genuinely county-dependent, which is why the report is built around the curve rather than around a recommendation. Where crop revenue per acre is high the sacrifice can be defended. Where it is low it cannot.

The honest boundaries of the work, in the order a reviewer would find them.

1. Crop revenue is small beside energy revenue. Electricity returns roughly 7,300 to 23,400 dollars per acre per year at benchmark prices. Crops return 225 to 4,560. Agrivoltaics improves the margin on commodity row crops; it does not transform the investment case, and the report should not be read as claiming otherwise.
2. Only 22 of 39 crops can be priced. Ten have no USDA commodity code at all. Five have a state price but no county yield. Two are grazing, which has no crop price by definition and is proxied by pastureland cash rent. These return blank rather than zero, and the page labels them.
3. Every retention factor traces to one systematic review. The review aggregates many field trials, and the individual trials and their geographies appear in the evidence notes, but the report has a single agronomic source and inherits whatever that source’s selection carries.
4. The market assumptions are mine. Installed cost and power price are national benchmarks from Berkeley Lab, exposed as adjustable parameters rather than presented as county-resolution facts. Payback is simple payback, excluding operations, degradation, tax credits and financing.
5. Coverage is uneven on the agricultural side. 685 counties have no usable yield-and-price pair. Of the 73 with no crop records at all, 62 are independent cities, boroughs and census areas with no farmland, which is a property of American administrative geography rather than a gap in the data.