How much of a solar forecast is skill? A two-site day-ahead study with public data

Day-ahead forecasts of photovoltaic (PV) output are usually reported as a single error number for one plant. This study asks how much of that number is skill, how much depends on the place, and where the skill runs out. It uses two public plants from the NREL PVDAQ archive in contrasting climates, an 893 kW fixed-tilt array in Arbuckle, California and a 4,738 kW single-axis tracker in Kersey, Colorado, and weather forecasts as they were issued a day ahead (GFS and ECMWF IFS, archived by Open-Meteo), with ERA5 reanalysis as a hindsight comparison. Every forecaster is scored on the same daytime hours with a monthly rolling origin: 6,009 hours over 549 days at Arbuckle and 3,078 hours over 298 days at Kersey. Against smart persistence (normalised RMSE 0.194 and 0.232 of AC rating), gradient-boosted trees on GFS inputs reach a skill of +0.25 (95% block-bootstrap interval +0.20 to +0.29) at Arbuckle and +0.22 (+0.14 to +0.29) at Kersey. A physical model with a two-parameter calibration gets most of the way there (+0.19 and +0.21); at Kersey the trees add nothing measurable over it. Hindsight weather from the reanalysis adds little at Arbuckle (+0.26) and somewhat more at Kersey (+0.29). A model trained on the other site keeps less than half of the gain at Arbuckle (+0.10) and about three quarters at Kersey (+0.16). Nominal 80% intervals cover 59% and 65% of hours. The worst tenth of days carries about a quarter of all absolute error. Read one by one, many of the largest misses are not errors of the weather forecast at all: a plant that produced nothing on two clear days and an inverter offline for a week; a summer of clear, dry days at reduced output may be equipment, soiling or smoke. Days on which the forecast was wrong and the reanalysis right are a minority of the large misses. All data are public, all code is released, and the whole study reruns with one command.