Year-one production matched the model. By year three the plant is running 9% under, nothing has obviously broken, and nobody can point at a cause. This is a common situation and it usually has an unglamorous explanation.
Here is the order I would investigate, cheapest and most likely first.
Start here, always. A pyranometer that has not been cleaned or recalibrated will drift. If it reads low, your PR looks artificially high and you will not notice a real problem. If it reads high, you will chase a shortfall that does not exist.
Check calibration date. Check the sensor is physically clean and unshaded. Cross-reference against a satellite irradiance dataset for your location – several are freely available and adequate for a sanity check.
Soiling in year one is light dust. By year three, in the wrong environment, you may have a bonded layer that ordinary rain no longer removes – cement dust, agricultural residue, salt film near the coast, or bird droppings that have baked on.
The tell: production recovers noticeably after heavy rain but never returns to year-one levels. That gap is the permanent fraction.
Also check the lower edge of the modules specifically. On low-tilt arrays a band of dirt accumulates along the bottom frame edge where water pools and evaporates. Because cells are wired in series, a soiled band across the bottom of every module can suppress the whole string far more than the affected area suggests.
Shading analysis is done once, at design. Trees keep growing. Neighbours build. Your own operations team installs a new extraction stack on the roof without anyone connecting it to the solar asset.
Walk the site at 9 a.m. and again at 4 p.m. in winter, when the sun is lowest and shadows longest. That is when new obstructions reveal themselves. Midday in June shows nothing.
This one is slow, invisible from the ground, and genuinely dangerous. DC connectors develop resistance through thermal cycling, moisture ingress and, commonly, mismatched-brand mating. Resistance produces heat, heat accelerates degradation.
The symptom is a string producing slightly less than its neighbours – often only 3-5% down, easily dismissed as noise. Thermal imaging under load finds these quickly. If you have never done an IR survey and the plant is past year two, this is the highest-value inspection available to you.
Inverters throttle when they overheat. Filters clog, fans fail, ventilation gets blocked by stored materials in a plant room nobody inspects. The plant keeps running, so nothing alarms, but peak-hour output is capped.
The signature is unmistakable once you look: a flat top on the production curve on hot days, at a level below rated output, that is absent on cool days.
Real, but slower than people assume. Modern modules degrade 0.4-0.55% a year after a first-year settling of 1-2%. By year three you should expect roughly 2.5-3.5% total. If your shortfall is 9%, degradation explains perhaps a third of it at most.
Faster degradation points to specific mechanisms – PID, LID, backsheet failure, cell cracking from bad handling during installation. These show up as distinct patterns in string data or in electroluminescence imaging, not as a uniform gentle decline.
The uncomfortable possibility. If the original yield model used 1% soiling in a dusty industrial setting, 100% availability, and no allowance for the shading that clearly exists, the plant is not underperforming – the model was wrong and everyone has been comparing against fiction.
You find this by rebuilding the model with honest assumptions and seeing whether the plant matches. It is an awkward conversation, particularly if the model came with a performance guarantee attached, but it is better than replacing hardware that is working correctly.
Most of that costs very little and answers the question. The instinct to start by replacing inverters is expensive and usually premature.