How often should a pyranometer be recalibrated?

Reference-grade pyranometers are typically recalibrated on a periodic schedule traceable to a recognised standard, with the interval set by the manufacturer's specification and the monitoring class you are maintaining. The operational principle matters more than the exact figure: recalibration should be a scheduled activity on a calendar, not a reaction to suspected error. A sensor recalibrated only when someone doubts it has already been drifting undetected up to that point.

Can you detect sensor drift without a reference instrument?

In most cases, yes. Your array itself acts as a large irradiance sensor, so a trending ratio of array output to measured irradiance on clear-sky intervals — with soiling and temperature accounted for — points to the sensor when the array has not changed. Comparing multiple on-site sensors against each other, and against a clear-sky model, triangulates the finding. A second instrument confirms it but is rarely required to detect it.

Does soiling on the sensor itself affect performance ratio?

Yes, and in a particularly misleading way. If the irradiance sensor soils at the same rate as the array, the two effects partly cancel and PR looks stable while real energy falls. If the sensor soils faster or slower than the modules, it introduces a bias in either direction. Because the sensor is small and easy to clean, its soiling state often differs from the array's — which is why sensor cleaning and inspection belong in the O&M routine, not just module washing.

What does IEC 61724 require for Class A monitoring?

IEC 61724-1 defines monitoring system classes — A, B and C — with Class A being the most stringent. Class A carries the tightest requirements for measured parameters, sensor type and accuracy, including reference-grade irradiance measurement with traceable calibration. Operators should confirm the current edition and its specific clauses when specifying a system, as standards are periodically revised.

Should historical reports be restated after a drift finding?

If the drift materially changes a figure someone relied on — a settlement, a guarantee test, an investor report — the responsible action is to restate it with a dated, visible correction note. For internal trending, re-baselining the affected period against a corrected reference (reconstructed from array and clear-sky data) restores the integrity of your history. Quiet edits discovered later damage credibility more than the original error did.

Irradiance Sensor Drift: Why Your Performance Baseline Is Quietly Wrong

Ellume Engineering

Irradiance sensors drift as they age and soil, and because almost every solar KPI is calculated against measured irradiance, that drift silently corrupts performance ratio, availability and loss figures. A sensor reading a few per cent low makes a healthy plant look like it is outperforming — and hides real underperformance. Detecting drift rarely needs a second instrument; it needs the right comparison.

Key Takeaways

→Nearly every solar KPI is computed against measured plane-of-array irradiance, so the least-audited instrument on site sets the denominator for performance ratio, availability and loss attribution alike.
→Drift is dangerous because nothing looks broken: the plant stays online, the dashboard stays green, and a sensor reading low makes an underperforming plant present as healthy.
→Pyranometers and reference cells drift for different reasons — dome soiling and thermopile ageing versus PV-cell degradation and spectral sensitivity — and confusing them leads to the wrong correction.
→A 3% instrument error is not a 3% reporting error: it lands on the denominator of almost every calculation, and it points in the direction least likely to trigger an investigation.
→Three comparisons using data you already hold — array-to-sensor ratio, sensor-to-sensor spread, and clear-sky peaks — triangulate drift without buying a single new instrument.

Here is a quietly uncomfortable idea to start the day with: the single most important number on your solar plant is probably measured by the least-audited instrument on site. Performance ratio, availability against a generation target, every megawatt-hour you have ever attributed to weather versus hardware — nearly all of it is computed against measured plane-of-array irradiance. The pyranometer or reference cell feeding that number is a small sensor bolted to a module rail, exposed to the same sun, dust, heat and weather as the array it is judging. It ages. And when it ages, it does not fail loudly. It drifts. A drifting irradiance sensor is dangerous precisely because nothing about it looks broken. The plant stays online. The dashboard stays green. The monthly report reconciles. But the reference the whole report is measured against has quietly moved, and every downstream figure has moved with it.

A sensor that reads low makes an underperforming plant look healthy. A sensor that reads high manufactures a problem that was never there.

What sensor drift actually looks like in the data

Drift is gradual and directional, which is exactly what makes it hard to spot. A step change gets noticed — someone asks why the number jumped. A slow slide of a fraction of a per cent per month does not trigger anyone's attention, because on any given day the reading looks entirely plausible. The tell is not in the irradiance signal on its own. It is in the relationship between irradiance and everything else. When a sensor drifts low, your calculated performance ratio drifts high, because PR is energy divided by expected energy and expected energy is built from that irradiance reading. A plant whose PR is creeping upward month over month, with no maintenance to explain it, is not necessarily improving. It may be measuring itself against a shrinking yardstick. The same logic runs in reverse. A sensor reading high — often after a recalibration that over-corrects, or a reference cell degrading differently from the modules it represents — pushes PR down and generates a phantom underperformance investigation that finds nothing, because there was nothing to find.

Pyranometer versus reference cell: two instruments, two failure modes

The two common plane-of-array sensors drift for different reasons, and confusing them leads to the wrong correction. A thermopile pyranometer measures broadband solar radiation across the full spectrum. It is the reference-grade choice, but it responds slowly, is sensitive to soiling on its dome, and needs periodic recalibration against a traceable standard. Its drift tends to come from soiling, dome degradation and the slow ageing of the thermopile itself. A reference cell is a small photovoltaic device, usually the same cell technology as the array. It responds fast and shares the spectral behaviour of the modules — which is its advantage — but it degrades the way a PV cell degrades, and a reference cell that ages faster or slower than the array it represents introduces a bias that grows over time. It is also sensitive to spectral conditions and temperature in ways a pyranometer is not. The practical consequence: a reference cell and the modules it stands in for can both be perfectly functional and still disagree, because they are degrading along slightly different curves. That divergence is not a fault in either device. It is a measurement gap that widens quietly with age.

How much error a small drift injects

This is where it stops being academic. Consider an illustrative case — a 120 MW single-axis-tracked plant in ERCOT, with a plane-of-array reference cell that has drifted 3 per cent low over four years without recalibration. The numbers below are illustrative, chosen to show the mechanism rather than to report a specific site. Because expected energy is computed from the low irradiance reading, expected energy is understated by roughly the same proportion. The plant's calculated performance ratio is inflated accordingly. A true PR of, say, 82 per cent presents as roughly 84–85 per cent. That two-to-three-point lift is enough to do real damage:

•It masks genuine loss. A string or tracker problem costing two points of PR is exactly cancelled by two points of phantom gain from the drift. The plant looks flat when it is actually losing.
•It corrupts availability-linked settlements. Where an O&M guarantee or an offtake arrangement references a performance figure, a biased reference cell moves money.
•It compounds in year-over-year comparisons. Each year is measured against a slightly more degraded sensor, so a fleet can decline steadily while every annual report shows it holding or improving.

The uncomfortable summary: a three per cent instrument error is not a three per cent reporting error. It is a three per cent error applied to the denominator of almost every performance calculation you make, and it points in the direction most likely to keep you from investigating.

Detecting drift without a second instrument

The instinct is to install a second, better sensor and compare. Sometimes that is warranted. But most drift can be caught from data you already collect, because a well-behaved plant contains its own redundancy.

1.Compare the sensor against the array it measures. Your modules are, collectively, a very large irradiance sensor. On clear-sky intervals, with soiling and temperature accounted for, the ratio of array output to measured irradiance should be stable over time. If that ratio trends while the array itself has not been touched, the sensor is the most likely thing that moved. This is the single most useful check, and it uses only inverter data and the irradiance reading you already have.
2.Compare multiple sensors against each other. Most utility-scale plants carry more than one irradiance sensor. They will never read identically — position, soiling and type differ — but the spread between them should be stable. A widening or narrowing gap between two sensors that used to track is a strong signal that one has moved.
3.Compare against a clear-sky model. On genuinely clear intervals, physics gives you an expected plane-of-array irradiance from the sun's position and a clear-sky model. Real readings sit below the model by a soiling-and-conditions margin, but the pattern of clear-sky peaks over months is revealing. A steady decline in clear-sky-peak readings, with no change in local climate, is consistent with soiling or drift on the sensor rather than a change in the resource.

None of these three is conclusive on its own. Together they triangulate: if the array-to-sensor ratio, the sensor-to-sensor spread and the clear-sky comparison all point the same way, you have drift, and you have it without buying a single new instrument.

Calibration intervals that actually hold in the field

The standards give you a framework here rather than a single magic number. IEC 61724-1, the standard governing PV system performance monitoring, defines monitoring classes — Class A being the highest — with associated requirements for sensor type, accuracy and calibration. For a Class A system, irradiance measurement is expected to use reference-grade sensors with periodic recalibration traceable to a recognised standard. The important operational point is not the exact interval but the discipline: recalibration is a scheduled maintenance activity, not an as-needed reaction. A sensor recalibrated only when someone suspects a problem has, by definition, been drifting undetected until the moment of suspicion. Treat irradiance sensors the way you treat revenue meters — on a calendar, with traceable records — rather than the way most sites treat them, which is to install them at commissioning and never think about them again.

What to do with historical data once you find drift

This is the part that separates a mature operation from a nervous one. When you discover a sensor has been drifting, you now know that a span of historical reports was computed against a biased reference. The question is what to do about the numbers you have already published. The honest answer is uncomfortable but correct: if the drift materially changes a figure that someone relied on — a settlement, a guarantee test, a board report — the responsible move is to restate it with a dated correction note, not to quietly move on. A visible, dated correction buys back more credibility than silence preserves. Silent edits, discovered later, read far worse than the original error. For internal trending, the better path is often to re-baseline: reconstruct what the irradiance should have read using the array-based and clear-sky methods above, and recompute the affected KPIs against the corrected reference. This is exactly the case for a performance reference built from physics rather than from a single instrument — a reference that cannot drift, because it was never a measurement in the first place.

Frequently Asked Questions

How often should a pyranometer be recalibrated?: Reference-grade pyranometers are typically recalibrated on a periodic schedule traceable to a recognised standard, with the interval set by the manufacturer's specification and the monitoring class you are maintaining. The operational principle matters more than the exact figure: recalibration should be a scheduled activity on a calendar, not a reaction to suspected error. A sensor recalibrated only when someone doubts it has already been drifting undetected up to that point.
Can you detect sensor drift without a reference instrument?: In most cases, yes. Your array itself acts as a large irradiance sensor, so a trending ratio of array output to measured irradiance on clear-sky intervals — with soiling and temperature accounted for — points to the sensor when the array has not changed. Comparing multiple on-site sensors against each other, and against a clear-sky model, triangulates the finding. A second instrument confirms it but is rarely required to detect it.
Does soiling on the sensor itself affect performance ratio?: Yes, and in a particularly misleading way. If the irradiance sensor soils at the same rate as the array, the two effects partly cancel and PR looks stable while real energy falls. If the sensor soils faster or slower than the modules, it introduces a bias in either direction. Because the sensor is small and easy to clean, its soiling state often differs from the array's — which is why sensor cleaning and inspection belong in the O&M routine, not just module washing.
What does IEC 61724 require for Class A monitoring?: IEC 61724-1 defines monitoring system classes — A, B and C — with Class A being the most stringent. Class A carries the tightest requirements for measured parameters, sensor type and accuracy, including reference-grade irradiance measurement with traceable calibration. Operators should confirm the current edition and its specific clauses when specifying a system, as standards are periodically revised.
Should historical reports be restated after a drift finding?: If the drift materially changes a figure someone relied on — a settlement, a guarantee test, an investor report — the responsible action is to restate it with a dated, visible correction note. For internal trending, re-baselining the affected period against a corrected reference (reconstructed from array and clear-sky data) restores the integrity of your history. Quiet edits discovered later damage credibility more than the original error did.

Where this leaves you

The performance number at the centre of your operation is only as trustworthy as the instrument it is measured against — and that instrument is drifting whether or not anyone is watching. The good news is that catching it rarely requires new hardware, only the right comparisons against data you already hold. The better news is that the whole problem dissolves when your reference comes from physics rather than from a single ageing sensor.