The yield study in your loan file was written before the plant existed. It predicted what 20 MWp of modules would produce in a typical year of weather. Four years later the plant has meter readings, sensor logs, outage records and cleaning dates. A refinancing lender can read all of it, so the old forecast is no longer the best number available.
This guide shows how the yield is re-run at refinancing. It covers why the P50 moves and how measured data is tied to a long-term irradiance record. It also covers which uncertainty components shrink, how degradation is read from the record, and what the new lender’s reviewer checks. The worked example uses illustrative numbers and says so. The guide is part of our solar engineering hub.
Quick answer. At refinancing, a P90 built on the plant’s own meter and weather data can replace the pre-construction P90. The re-run calibrates the model to measured generation, corrects it to long-term irradiance, and measures degradation and availability from the record. Operating history shrinks model and measurement uncertainty, but not future weather. In the illustrative case below, one-year uncertainty falls from about 6.0% to 3.8%.
Why does solar plant refinancing re-base the yield?
At financial close the forecast came from a model run on long-term weather data. Each modelling step carried an assigned uncertainty. Those were combined into a P50, the annual energy exceeded in half of years, and a P90, exceeded in nine years of ten. Our guide to P50, P90 and P99 yield reports covers the formula and the pre-construction uncertainty stack. This guide does not repeat it.
Operation changes three things.
- Energy is measured. The billing meter records what the plant actually delivered, year by year.
- Model inputs can be checked. The plant’s own sensors show the irradiance and temperature the modules saw. The model can run on real weather instead of a typical year.
- Assumptions become observations. Degradation, availability and soiling were assumed at close. They are now in the data.
The IEA PVPS Task 13 report on yield uncertainty, IEA-PVPS T13-12:2018, splits a forecast into three steps. The first is the yield in a reference period. The second is the change over the prediction period, such as degradation. The third is year-to-year weather variability. Operating data tests the first two steps directly. It cannot test the third.
Advisers describe the same logic. DNV describes its operational energy assessment as a post-construction forecast based on the site’s operational data. It gives P50 and P90 values and offers uncertainty analysis to help with refinancing or budgeting. DNV’s page is written mainly around wind projects, so read it as a description of the approach, not a solar method. In December 2025 DNV also described its role as technical due diligence adviser on a €2 billion refinancing of 50 PV plants. Its stated scope included an energy assessment and performance reviews of operating plants.
Your lender sets its own scope. Nothing here says what your term sheet requires.
How do you combine measured data with long-term irradiance?
A plant has three to five years of data. A satellite or reanalysis record has twenty or more. The task is to let the short, accurate record correct the long, approximate one. Wind engineers call this measure-correlate-predict. Solar resource specialists call it site adaptation. Polo and co-authors, in a 2016 survey in Solar Energy, describe it as using short-term on-site ground measurements to correct systematic bias. The data being corrected is the long-term satellite-derived and model-derived radiation record.
The sequence below is how a yield engineer typically builds the re-run. It is our working order, drawn from the sources cited in each step.
- Audit the plant’s data. Check sensor type, calibration dates, soiling on the sensors and drift. The IEA report notes that monitoring data can be validated against calibrated instruments installed temporarily for a defined period, then corrected. IEC 61724-1:2021 defines the classes of PV monitoring systems and dropped Class C. Our IEC 61724 explainer covers sensor specifications by class.
- Build the measured-period weather. Use the plant’s plane-of-array irradiance (POA) and module temperature. If the long-term dataset gives GHI, converting between the two adds transposition uncertainty, and the study should say so.
- Adapt the long-term record. Compare the satellite series with the plant sensors over the overlap period, derive a correction, and apply it to the full record. Solargis publishes its own guidance in its site adaptation knowledge base. It recommends at least 12 months of quality-controlled ground data and calls 24 months or more optimal. It does not recommend 3 to 6 months. It also says ground data must be more certain than the satellite model, or adaptation cannot improve the record. That is vendor guidance, not a standard. For the dataset choice itself, see our comparison of Meteonorm, Solargis and NSRDB.
- Calibrate the model on measured weather. Run the PVsyst model with the measured weather for the operating period and compare it with metered energy. In the vocabulary of IEC TS 61724-3:2016, the ratio of measured to expected energy is the energy performance index. The specification separates two versions. The in-service index leaves out times when inverters are offline. The all-in index includes them. Any difference that remains after known causes are removed is model bias, and it is carried forward.
- Re-run the typical year. Run the calibrated model on the adapted long-term series. That gives the new P50 before degradation and availability are applied going forward.
- Rebuild the uncertainty and the P90. Each component is re-assessed, as the next section shows.
Step 3 matters because the measured years are not typical by default. NREL researchers found that typical meteorological year (TMY) files on average overpredict PV production against the long-term mean in the contiguous United States. They also found that no single year works as a representative resource year for all regions. A plant that happened to run through four sunny years looks better than it is. One that ran through four weak years looks worse. The long-term correction removes that luck.
If your plant needs its PVsyst model re-run against measured weather, our PVsyst energy simulation reports page describes the report type. The scope of a re-run is agreed at scoping.
Which uncertainty components shrink with operating history?
Sandia National Laboratories splits P90 uncertainty into two kinds, in a 2024 PVPMC workshop presentation by Clifford Hansen. Aleatory uncertainty is inherent and cannot practically be reduced. Future weather is the example. Epistemic uncertainty is state of knowledge, and better measurements, models or data can reduce it in principle.
Operating history attacks the epistemic part. It cannot touch the aleatory part. That one sentence is the reason a refinanced P90 does not collapse toward the P50.
The table below shows the effect on a one-year P90 for the illustrative 20 MWp plant used in the next section. The standard deviations (σ, as a percentage of P50) are chosen for the example. They are not taken from a source or from any real plant.
| Component | Before, at close (σ, %) | After 4 years of data (σ, %) | Why it moves |
|---|---|---|---|
| Long-term irradiance | 4.0 | 2.0 | Satellite-only record replaced by one adapted to four years of validated plant sensor data |
| Inter-annual variability | 2.5 | 2.5 | Next year’s weather is unknown, whatever the history shows |
| Measurement (meter and sensors) | not applicable | 1.0 | New component: the record is only as good as its instruments |
| PV model | 3.0 | 1.5 | Model calibrated on four years of measured energy and weather |
| Degradation | 1.0 | 0.5 | Assumed rate replaced by a measured rate with an interval |
| Availability | 1.5 | 0.8 | Outage history replaces an assumption |
| Soiling | 1.0 | 0.5 | Cleaning log and measured soiling replace a modelled loss |
| Combined (root sum of squares) | 6.0 | 3.8 | |
| P90 haircut (1.2816 × σ) | 7.6% | 4.9% |
The combined figures come from squaring each value, adding them and taking the square root. Before: 16 + 6.25 + 0 + 9 + 1 + 2.25 + 1 = 35.5, and the square root is 5.96. After: 4 + 6.25 + 1 + 2.25 + 0.25 + 0.64 + 0.25 = 14.64, and the square root is 3.83. Adding in quadrature assumes the components are independent. The IEA report notes that independence is usually assumed. It is a simplification.
Three notes on the table.
Measurement is a new line, and it can be small or large. NREL’s uncertainty report, NREL/TP-7A40-84993 (revised March 2024), cites an analysis of research pyranometers at one site. It put the effect on annual energy at a standard deviation of 0.35% to 0.6%. Hourly pyranometer uncertainty is often reported at 3% to 20% at 95% confidence. The instruments in that analysis were maintained daily and the report regards their data as among the best achievable. An unmaintained plant sensor is a different matter. The IEA report gives billing meter uncertainty as 0.5% or 0.2%.
Availability starts from a wide assumption. The same NREL report lists a default triangular distribution for inverter availability in the System Advisor Model. It is wide because nobody knows at close when an inverter will fail or when maintenance will fall. A plant that logged its outages has replaced that spread with a record.
Inter-annual variability stays. For a one-year P90 it stays at full size. For a P90 of the ten-year average it shrinks, because high and low irradiance years offset. In the IEA report, individual years carry the full inter-annual spread, while cumulative yields lose most of it within a few years. Ask your lender which one its debt sizing uses.
How do you measure degradation from plant data?
At close, degradation was a number from a datasheet or a study. The experts in the IEA survey disagree about it. One uses a first-year drop of 1% to 2% and then 0.3% to 0.5% a year. Another uses 0.25% a year for crystalline silicon. Plant data replaces the argument with a measurement.
The older method is a linear regression through monthly performance ratio values. Seasonality hurts it. The IEA report cites three years of data as the minimum for reasonable uncertainty with that approach, and notes that low-seasonality metrics work better.
The method now favoured is year-on-year (YoY). It compares the plant’s normalised performance at one point in time with the same point a year earlier. It then takes the centre of the resulting distribution of rates. Jordan and co-authors (IEEE Journal of Photovoltaics, 2018) report that most published methods are biased by clipping, soiling, temporary outages, seasonality and sensor degradation. They propose modelled clear-sky irradiance with a YoY calculation. In their comparison across 486 systems, it had the lowest uncertainty of the methods tested.
NREL’s open-source library RdTools implements it and calls YoY its preferred method. The degradation workflow in its documentation runs in five steps.
- Normalise measured power by expected power from a model that uses the plant’s irradiance and cell temperature.
- Filter out invalid, low-irradiance and clipped points.
- Aggregate to daily values, weighted by irradiance.
- Remove anomalous daily values.
- Run the YoY calculation, which returns a rate, an interval and, if requested, an exceedance value such as P95.
On its public sample data the example notebook reports a rate of about 0.41% a year, with a 68.2% interval of 0.30% to 0.63% a year. The documentation also states two limits that matter for a refinancing file. Automatic outage detection is not yet included, so outage filters must be built for the plant. And because sensors drift, the notebook recommends checking sensor-based results against a satellite-based analysis.
Four practical points follow.
- State how the first year is treated. Initial light-induced drop sits in the first year of the record. Report the rate with and without it.
- Carry the interval forward. Take a rate of 0.8% a year with an interval of 0.2% a year. Extrapolated 2.5 years beyond the middle of the record, it adds about 0.5% to the uncertainty of a year-5 forecast. That is the 0.5 in the table above.
- Use context, not a substitute. Jordan and Kurtz compiled nearly 2,000 published rates in an NREL review and found a median of 0.5% a year. The spread around that median is the reason a plant-specific measurement is worth having.
- Do not confuse model degradation with measured degradation. The model assumptions are explained in our guide to PVsyst degradation modelling. The degradation rate entry has the definition.
How do you treat availability, curtailment and soiling in the record?
Each of these three losses was a guess at close. In the data they can be separated and labelled.
Availability. IEC TS 61724-3 treats times when the plant is not operating as an availability matter. The energy performance index covers the times when it is. RdTools also lists availability analysis among its applications. Tag every outage with a cause and a duration. Plant-side causes, such as inverter faults and tripped breakers, belong to availability. Grid-side causes, such as feeder outages or backdown instructions, belong to curtailment. Keep them in separate columns. The forward argument for each is different, and your PPA decides how grid-side energy is treated.
Clipping and shading. DNV’s article on SCADA data (9 June 2020) describes a UK solar park investigation. Inverter-level data identified when clipping occurred and let the team calculate the energy lost. String-level data separated inter-row shading from external shading. The article also observes that periodic reviews are often limited to monthly performance ratio. A monthly figure hides these effects, and the re-run should not.
Soiling. Compare the cleaning log with measured soiling loss where a soiling station exists. The forecast must assume that the cleaning regime continues, or state what changes. Regional behaviour is covered in our India soiling guide.
What does a re-based P90 look like for a 20 MWp plant?
The numbers below are illustrative. They describe an invented plant. They are not a forecast for any real project, and they are not a Heaven Designs result.
Assumptions
- A 20 MWp ground-mount plant in India, four years into operation.
- A flat tariff of ₹3.00 per kWh (illustrative).
- Original year-1 P50 of 1,650 kWh/kWp, with 0.7% a year degradation assumed.
- Original year-5 P50: 1,650 × (1 − 0.007 × 4) = 1,604 kWh/kWp, which is 32.08 GWh a year.
- The operating study finds a weather-normalised year-5 P50 of 1,570 kWh/kWp, which is 2.1% lower.
- The invented attribution of that 2.1%: availability 0.6 points lower than assumed and soiling 0.9 points higher. Measured degradation of 0.8% a year against 0.7% assumed adds 0.4 points by year 5. The last 0.2 points are other causes.
- Uncertainty as in the table above: σ of 6.0% before and 3.8% after.
Result
| Basis | Year-5 P50 (kWh/kWp) | P50 revenue (₹ crore) | Year-5 P90 (kWh/kWp) | P90 revenue (₹ crore) |
|---|---|---|---|---|
| Original forecast | 1,604 | 9.62 | 1,481 | 8.89 |
| Re-based on operating data | 1,570 | 9.42 | 1,493 | 8.96 |
The arithmetic: the original P90 is 1,603.8 × (1 − 0.0764) = 1,481 kWh/kWp, or 29.63 GWh, which at ₹3.00 is ₹8.89 crore. The re-based P90 is 1,570 × (1 − 0.0490) = 1,493 kWh/kWp, or 29.86 GWh, which is ₹8.96 crore.
The P50 fell by ₹0.20 crore a year. The P90 rose by ₹0.07 crore a year. The shortfall was real, but the narrower uncertainty more than made up for it at the P90 level.
That result depends on the size of the shortfall. The break-even is the ratio of the two P90 factors: 0.924 divided by 0.951 is 0.971. A measured shortfall of up to about 2.9% leaves the new P90 higher than the old one. A larger shortfall pushes it lower. At a 5% shortfall, the new P90 is 1,449 kWh/kWp, or ₹8.69 crore, about ₹0.20 crore below the original P90.
A lender sizes debt on P90 revenue after costs. A re-based P90 is therefore a financing input, not a technical footnote. Whether the number helps or hurts depends on what the plant did, and on how well the study can evidence it.
What does the new lender’s reviewer check?
The reviewer works for the lender, and the lender chooses and sets terms for it. Our comparison of the owner’s engineer and the lender’s engineer explains the roles. The list below is built from the sources above, and it is the set of questions a careful reviewer can ask of a re-run. It is not a lender’s published checklist.
- Data provenance. Which sensors, of what class, calibrated when, cleaned when? Is there a calibration certificate for each?
- Completeness. What fraction of intervals is missing? How are gaps filled, and are filled values flagged?
- Reconciliation to the meter. Does summed plant energy match billing meter statements, such as joint meter reading statements where the PPA uses them?
- Outage and curtailment log. Is every energy-loss estimate tied to a cause code and a time window?
- Model changes. What was recalibrated relative to the original model, and why? IEC TS 61724-3 expects the model to stay fixed within a test, so a revision should be a visible, dated step.
- Long-term dataset. Which source, which years, how long is the overlap with plant sensors, and how was bias corrected? Is a second dataset used as a cross-check?
- Degradation. Which method, which filters, which interval, how was the first year treated, and was sensor drift checked?
- Uncertainty budget. Is each component listed with a source? Does every component that fell have a stated reason?
- Forward assumptions. Do cleaning, operation and maintenance, inverter replacement and any repowering plans match the history? If equipment changes, the model must change. See our guide to solar plant repowering in India.
- Bridge to the original. Is the difference between the old P50 and the new P50 explained, line by line?
Items 1 to 3 and 5 overlap with pre-submission checks for any yield report. Our guide to validating a PVsyst report before a lender sees it covers those.
What data should you gather first?
Collect the following before anyone runs a model.
- Inverter-level time series at the finest interval stored, plus plane-of-array irradiance, module temperature and ambient temperature.
- Sensor calibration certificates and cleaning records.
- The original yield study, the PVsyst project file and any as-built changes.
- Operation and maintenance logs, including outages, inverter and module replacements, and cleaning dates.
- Billing meter statements and any grid outage or backdown records.
- The handover performance ratio test, if one was run. It is a useful year-zero baseline. See our guide to the performance ratio test at handover.
- The engineering documents already in the loan file. See solar project loan documents in India for the list.
Bring as many full years as the plant has. A short record is usable, but it widens the intervals, as the sources above show.
Where does Heaven Designs fit?
Heaven Designs is a solar design, engineering and consultancy firm, founded in 2020 in Surat (company-stated). We are not a lender’s engineer or an independent engineer, and a refinancing lender picks its own reviewer. A re-run commissioned by the owner is an owner-side document. The reviewer will test it.
Our reports page lists PVsyst energy simulation reports with loss breakdown, performance ratio analysis and shading analysis. It says we tailor reports to lender and client requirements. If you want the simulation re-run against your plant’s measured weather, say so when you contact us. Scope and turnaround are agreed at scoping.
FAQ
Does refinancing always require a new yield study?
No source we reviewed sets a universal rule. The lender’s term sheet or technical due diligence scope decides. Some lenders review operating performance as part of due diligence, as DNV’s Sonnedix announcement describes. Ask your lender what it needs before you commission any work.
How many years of operating data do you need?
It depends on the task. Solargis recommends at least 12 months of quality-controlled ground data for site adaptation. IEC TS 61724-3 prefers one year for an energy evaluation. The IEA report cites three years as the minimum for a linear-regression degradation estimate. A shorter record still helps, but its intervals are wider.
Can the P90 go up after refinancing?
Yes, if the measured shortfall against the original P50 is smaller than the drop in uncertainty can absorb. In the illustrative example the break-even is a shortfall of about 2.9%. If the plant under-performed by more than the break-even, the P90 falls.
Does the owner’s study replace the lender’s reviewer?
No. The lender chooses its own reviewer and sets its terms. The owner’s study gives that reviewer a documented basis to test. A firm that prepared the owner’s study should not also act as the lender’s reviewer on the same plant.
What if the plant’s irradiance sensors are poor or missing?
The IEA report describes validating monitoring data against calibrated instruments installed temporarily, then correcting it. Without usable sensors, the study falls back on satellite data alone, and the resource uncertainty stays at the pre-construction level. The P90 benefit shrinks accordingly.
Next step
List what the plant has: years of data, sensor calibration records and the original yield study. Then ask your lender what a refinancing review needs. For a PVsyst re-run against measured weather, see our PVsyst energy simulation reports page, and send us the plant details so scope can be agreed. We reply within 1 business day.
This guide was reviewed against the cited sources on 31 October 2026. Standards editions, documentation and lender terms change, so recheck them before you rely on them. It is engineering guidance, not financial or legal advice. The worked example is illustrative, and no lender has agreed that any method described here will be accepted.