Late inverter starts cut 1.88% of annual output at one solar plant

At sunrise in a utility-scale solar plant in western São Paulo, irradiance could already support generation while some inverter units—the equipment expected to begin delivering power—were still waiting. Researchers from the University of Campinas (Unicamp) found that inverter startup delays resulted in annual energy losses equivalent to 1.88% of total generation, after analyzing 2023 operational data.

The team used sunrise as the reference point and compared expected startup behavior with the plant’s AC (alternating-current) active-power data. At the site, delays reached up to four hours after sunrise, even when irradiance levels were sufficient for power generation. The approach was designed to distinguish startup losses from other causes, including system unavailability, component mismatch and weather conditions.

The losses were concentrated at the beginning of the day. More than 55% of the annual losses linked to delayed startup occurred within the first 90 minutes after sunrise. They also changed with the seasons: the first half of the year accounted for 68.8% of the recorded losses in the Southern Hemisphere’s summer and autumn. The eight inverters did not behave alike, with monthly losses ranging from 0.14% to 5.02% of each inverter’s energy generation.

Concretely, operators could use sunrise times and existing AC power records to identify which inverter units start late, then focus operational interventions on the largest sources of lost generation. Plane-of-array, or POA, irradiance measurements are needed only for configuration and validation, reducing the instrumentation required for the analysis.

The result remains a case study from a single plant, not a universal loss rate for solar facilities. The Unicamp researchers said the 1.88% figure is comparable to losses typically considered in photovoltaic performance and availability assessments, and that their methodology can be applied to utility-scale plants using existing operational data.