Industries / Construction and property / Commercial and residential construction / Electrical contracting / Solar installation
Inspection and monitoring
This work tests a finished system before it is switched on, inspects it later, and watches its output to find faults, dirt and slow decline.
6
Milestones achieved, from 2004 to 2020.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2004
A shared record of output
Performance ratio is the energy a system delivers divided by what its panels' rated power would give under the sunlight that actually fell on them; it lets systems in different places be compared. Through Task 2 of the IEA's Photovoltaic Power Systems Programme, Ulrike Jahn and Wolfgang Nasse compared the performance ratios of 235 German installations and 133 elsewhere and reported in 2004 that newer German systems performed and held up better, thanks to more efficient inverters and less downtime. A 2012 Fraunhofer ISE study of about 100 German systems found performance ratios between about 70% and 90% in 2010, with a median of about 84%.
LessonThe 2004 paper noted a lack of long-term experience with how systems age "owing to the absence of monitoring programmes"; performance is only known where someone records it.
Credit
Ulrike Jahn and Wolfgang Nasse (IEA PVPS Task 2); Nils H. Reich and colleagues (Fraunhofer ISE)
2009
Tests before switch-on
IEC 62446, published in 2009 and replaced in 2016 by IEC 62446-1, sets the documents an installer hands to the owner and the tests that show a grid-connected system is safe and working. Every system gets the "Category 1" tests: polarity, the open-circuit voltage and short-circuit current of each string (a chain of panels wired in series), and the insulation resistance of the array, a check that current cannot leak to the frame or the ground. An extended "Category 2" set adds a current-voltage (I-V) curve for each string and an infrared camera scan. The 2016 edition widened the tests to cover larger and more complex systems.
Credit
International Electrotechnical Commission (IEC)
2010
Random checks on home systems
Australia's Clean Energy Regulator sends inspectors to a random sample of the small systems registered under its Small-scale Renewable Energy Scheme. From 2010 to 2018 it made 28,041 inspections and rated 3.3% of systems unsafe or potentially unsafe; the share fell to 1.7% among 2018 installations. In the 2025–26 year, of 3,129 randomly chosen systems installed since July 2024, 0.93% were unsafe and were shut down or made safe on the spot, and 23.11% were "substandard": safe to run but short of one or more clauses of the Australian standards. Installers with such results must return, fix the work and send proof to the accreditation body.
LessonA random sample shows the trend across a whole market; the unsafe share fell, but nearly one new system in four still misses the standard.
Credit
Clean Energy Regulator (Australia)
2012
A country's roofs scored
Jonathan Leloux, Luis Narvarte and David Trebosc analysed operating data from 6,868 home systems in France. In 2010 these produced a mean of 1,163 kWh for each kW of panels, with a mean performance ratio of 76%. In 2020 the same group, with Adrien Desportes, found faults in about 6,000 European systems by comparing each system's output with its neighbours' ("performance to peers"), a method that needs no sunlight sensor on site.
LessonOutput data shared by owners can audit a whole market.
Credit
Jonathan Leloux, Luis Narvarte, David Trebosc and Adrien Desportes
2019
Dust costs 3 to 4%
Soiling is the dust, pollen, bird droppings and other dirt that settles on panels. Klemens Ilse, Leonardo Micheli, Benjamin Figgis, Katja Lange and colleagues estimated that, even with cleaning schedules chosen to keep costs lowest, soiling cut the world's solar output by at least 3 to 4%, worth at least €3 to 5 billion a year, and that this could rise to 4 to 7%, more than €4 to 7 billion, by 2023.
Credit
Klemens Ilse, Leonardo Micheli, Benjamin W. Figgis, Katja Lange and co-authors
2020
Plants slip 1.3% a year
Mark Bolinger, Will Gorman and Dev Millstein of Lawrence Berkeley National Laboratory and Dirk Jordan of NREL compared the output of 411 large US solar plants, 21.1 GW of panels that started between 2007 and 2016, with what their sunlight should have given. The plants met expectations in their first year, then lost 1.3% (±0.2%) a year, against the 0.5% a year commonly assumed. The figure covers the whole plant, including dirt, other equipment and downtime, not the panels alone. A 2022 update over 31 GW found 1.2% a year, and Berkeley Lab's 2024 utility-scale report, covering plants built through 2023, put newer plants at 0.9% a year against 1.47% for older ones.
LessonPanel ageing explains only part of the decline; what a whole plant loses has to be measured at the plant.
Credit
Mark Bolinger, Will Gorman, Dev Millstein (Berkeley Lab); Dirk Jordan (NREL)
Watch for
More inspections, less loss
Raptor Maps, a US company that analyses solar inspection data, says in its 2026 Global Solar Report, covering 373 GW of US utility and commercial sites, that equipment problems cost an average 5.08% of power in 2025, and 4.46% at sites just commissioned. It says sites inspected by autonomous drones averaged 3%, and sites checked quarterly did 36% better than sites checked once a year.
Achieved whenan outside study compares sites matched for age, size, region and owner and still finds the gap; the report itself says larger, better-resourced sites perform better, so the sites that buy drones may simply be better run.
Claimed by
Raptor Maps, February 2026
Counting uncounted rooftops
HeraldX and Renewables First say in their June 2026 report "The Uncounted Gigawatts" that an AI model trained on Pakistani rooftop imagery found 177 MW of rooftop solar in one grid area of Lahore where the official estimate was 95 MW, and that independent estimates put Pakistan's distributed solar at about 38 GW, nearly twice the official record. Academic teams have also counted arrays from satellite images: Muhammad Kamran Lodhi and colleagues found more than 85,000 in Islamabad and Lahore in a 2025 conference paper, without comparing them to connection records.
Achieved whena distribution company or Pakistan's power regulator confirms the count by field survey or meter records, or the authors publish the model's accuracy on a hand-checked sample of rooftops.
Claimed by
HeraldX and Renewables First, June 2026
Open
01
Answer key for fault finders
A public set of output records from at least 200 home systems in which every fault period is dated and confirmed by an installer's job sheet, the owner's photos or a repair receipt: new shade from a tree, an inverter that keeps tripping, a dead string, dirt, snow, an export limit set by the utility. Published fault-finding methods are then scored on the same records.
The briefCredit · unclaimed02
Test the year-three drop
An open, repeatable analysis of US solar plants' output by year of age, adjusted for the sunlight each year, that tests whether losses jump between the second and third year of operation, when construction warranties often end, with the code published.
The briefCredit · unclaimed
Open 01 · the brief
Answer key for fault finders
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of output records from at least 200 home systems in which every fault period is dated and confirmed by an installer's job sheet, the owner's photos or a repair receipt: new shade from a tree, an inverter that keeps tripping, a dead string, dirt, snow, an export limit set by the utility. Published fault-finding methods are then scored on the same records.
Why it is still open
Methods for spotting faults in home systems have been tested mostly on data others cannot see: Leloux and colleagues' "performance to peers" on about 6,000 European systems, and a UNSW-led project on systems supplied by five New South Wales councils. The open sets with labelled faults found here cover a single home each, such as a 4 kW roof on Réunion island with labelled shading faults (Graillet and colleagues, 2023); SolarFCD (April 2026) pools 4,435 images, not output records, and the solar fault sets on Hugging Face are also images.
Built with what exists
PVOutput.org, where owners already share readings (81.8 million outputs recorded, by its own count); the open-source libraries pvlib (expected output), PVAnalytics (data cleaning and labelling) and RdTools (soiling and decline); the SunDown method, which compares neighbouring panels; an AI model that sorts job sheets and receipts into fixed fault categories (see the Note "A thousand questions, one pass").
First step
Ask 20 PVOutput contributors in one region to list every fault and repair since 2024 with its dates, and line each one up against their records.
Done looks like
200 homes and at least 300 confirmed fault periods in at least six categories, released under an open licence, with at least three methods (a peer comparison, a physical model, an AI model) scored on faults found, faults missed and false alarms.
Open 02 · the brief
Test the year-three drop
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open, repeatable analysis of US solar plants' output by year of age, adjusted for the sunlight each year, that tests whether losses jump between the second and third year of operation, when construction warranties often end, with the code published.
Why it is still open
Raptor Maps reported in February 2026 that average power loss in its data rose 55% between years two and three, which it ties to builders' warranties of one to two years; its data come from its own customers. Berkeley Lab's plant-level studies estimated performance at each age and found a near-straight decline, and its 2024 report splits newer from older plants, but none of them tests the step from year two to year three, and the 2022 study says its model cannot explain why some older plants improved.
Built with what exists
The US Energy Information Administration's public plant lists (Form EIA-860M) and monthly generation (Form EIA-923); pvlib to model what each plant's sunlight should give; Berkeley Lab's published method (age fixed effects, which compare each plant only with itself as it ages); an AI coding agent to build and document the pipeline.
First step
Download EIA-923 monthly generation for 2018 to 2025, pick 100 plants that started in 2019 to 2021, and plot their sunlight-adjusted output for years one to four.
Done looks like
Age curves with 95% confidence intervals for at least 1,000 US plants of 1 MW or more, through 2025, with code and data released, and a stated answer to whether the change from year two to year three differs from other one-year changes, set beside Raptor Maps' 55%.