Skip to content

Is the flat top of metered solar farms C, D, E, and F a fixed export cap or a panel-plane effect?

Four metered solar farms (anonymised as farms C, D, E, and F) show a flat top, meaning output that stops rising at a roughly constant level around midday on clear days. That level is lower in summer than in spring, so the flat top does not behave like a fixed export cap, a limit on grid export. We asked whether each farm's flat level, measured as power divided by capacity, stays constant across the months (as an export cap would) or changes with the season (as heat derating or the panels' tilt and azimuth could make it). We took the half-hours in which the modelled clear-sky irradiance on a horizontal plane exceeds 650 W/m², from the cleaned power series of the metered-solar-farm exploration (experiment 179). For each farm and month from April to August we then took the 90th percentile, across those half-hours, of power divided by capacity, which the page calls the flat level. No forecasting model was fitted.

The pre-registered prediction fails and the stop rule fires. The prediction was that at least 3 of the 4 farms keep a flat level that varies by under 3% across months, where the variation is the range of the monthly levels divided by their mean. The variation is 5.9% to 8.9%, so no farm is seasonally constant. The stop rule then says that the per-farm plant fit (follow-up 2, a fit of each farm's panel tilt, panel azimuth, direct-current (DC) capacity, and alternating-current (AC) capacity) must fit tilt and azimuth, not only a clip level. September has under 30 qualifying half-hours at every farm and is left out. The 650 W/m² threshold is the irradiance above which experiment 179's figure showed the flat top.

Figure 1: monthly flat level per farm, at the 90th and 99th percentiles of power over capacity

Numbers

  • Flat level, 90th percentile, range over mean, April to August, farms C to F: 7.2%, 8.9%, 5.9%, and 8.6%. Each farm's level is highest in April (1.04 to 1.05 times capacity) and lowest in July (0.96 to 0.98 times capacity).
  • Replication by years: even years alone give 5.0%, 8.7%, 6.4%, and 8.1%. Odd years alone give 9.8%, 13.2%, 5.4%, and 11.9%, and farm E's odd-year figure covers April to July only. Every farm varies by more than 3% in both halves, so the result of 0 of 4 constant farms is replicated.
  • 99th percentile (post hoc), range over mean: 1.4%, 5.8%, 2.7%, and 3.4%. Farms C and E would pass the 3% rule at this percentile. Farm F narrowly fails at 3.4%, and farm D fails at 5.8%. All four 99th-percentile levels still fall from April to August (for example farm F, 1.06 to 1.03 times capacity).
  • Morning against afternoon (90th percentile, same half-hours): the ratio is 0.99, 1.02, 0.99, and 1.00 for farms C to F, so the level does not depend on whether the sun is east or west of south. A south-facing array and a fixed cap would both give this ratio, so the ratio does not separate them.

Checks

  • More cloud in summer could lower the 90th percentile. The clear-sky filter uses the clear-sky model, not observed irradiance, so cloudy half-hours enter the 90th percentile. The 99th percentile is less exposed to cloud, and its range across months is smaller. Part of the 90th-percentile range may therefore be cloud, but the seasonal fall at the 99th percentile remains at all four farms, and at farm D its range (5.8%) is close to the 90th percentile's (8.9%).
  • Heat derating, panel tilt and azimuth, and a seasonal export cap are not separated, because no temperature, no panel tilt or azimuth, and no export-cap record was used.
  • Absolute levels are not meaningful, only their change across months. Power is divided by the effective capacity from studies.pv_dataset, which may itself be estimated from the output, so a level near 1.0 is partly by construction.
  • A monthly difference could come from sun elevation. The qualifying half-hours in April and August are near midday only, so month and sun elevation are confounded.
  • Audit: the audit rule for a result that contradicts its own prediction applies in form. The explanation that cloud lowers the 90th percentile was tested with the 99th percentile above, and the seasonal fall remains. No audit was run.

What the next experiment should answer

  • Does a per-farm physical plant fit (panel tilt, azimuth, DC capacity, AC capacity, and a heat-derating loss) reproduce the fall from April to August better than a clip level alone? The prediction is that it does; it fails if the clip-only fit does as well.
  • Does each farm's flat top coincide with the hours in which the network operator limited grid export (studies.export_cap identifies these hours)?