Why two solar production estimates differ and how to reconcile them
Two reports on the same plant disagree by 4 %. Before you ask which tool is wrong, look at the published evidence on how much disagreement is normal — because it is more than most people assume, and the largest single source is not the software.
This page is for the engineer who has to explain a discrepancy to a project buyer, an independent engineer or an investment committee. It is a procedure, not a defence of any tool. Heliosolve publishes it; the measurements cited are other people’s.
Four layers of disagreement, and none explains the others
| Layer | What was measured | Published figure |
|---|---|---|
| Between commercial tools | 7 vendors, fixed-tilt monofacial, sub-hourly, 1 year, weather given | Mean error in annual yield 2.5 % (15.4 kW lab system) rising to 6.0 % (14.5 MW utility plant) |
| Between people | 5–7 professional assessors, identical inputs, same plant | P50 from 1,095 to 1,405 kWh/kWp; σ ≈ 9.7 % on complex terrain |
| Inside specific physics | Rear-side irradiance, thermal models, sub-hourly clipping, backtracking on terrain | Rear rRMSE ≈21–23 % vs ≈7 % front; clipping under-counted by up to 1.5 %; ≈2 %/yr from overridden backtracking |
| Against the operating fleet | >200 utility-scale plants, weather-adjusted | 7–13 % below the original P50 for plants built after 2015 |
Sources and caveats for each figure are below. The point of the table is the structure: these are four independent gaps. Closing the tool gap does not close the assessor gap, and neither of them explains the fleet gap.
Layer 1 — how far apart the tools are on the easy case
The reference measurement is published as a feature review of the tools using blind assessment: Sandia’s PV Performance Modeling Collaborative gave vendors the site and weather data, collected the predictions the vendors themselves produced, and only then compared against measurements. Seven commercial tools took part — 3E SynaptiQ, PlantPredict, PVsyst, RatedPower, SAM, SolarFarmer and Solargis Evaluate — on two systems: a 15.4 kW laboratory array in Albuquerque and a 14.5 MW plant in Germany, at sub-hourly resolution over one year.
Reported result: “deviations ranging from mean error in annual yield up to 2.5 % for laboratory systems, increasing to 6.0 % for utility-scale installations”, attributed to “combined user decisions and inherent software behavior” (Deville et al., Solar Energy 304, 114207, 2026; we read the abstract and publisher metadata, not the full text).
Two things this does not say, and both matter because they are widely misquoted:
- This study did not test bifacial modules, trackers or terrain. Both systems were fixed-tilt, monofacial and south-facing. The increase from 2.5 % to 6.0 % is with scale, not with optical complexity. There is another blind exercise that did cover tracking and bifacial, and it sits in layer 2.
- It does not rank the tools. Vendor-attributed residuals are not something you can lift from a press summary.
So: on the simplest configuration in the industry, with the weather file handed over, the spread in annual energy is a few per cent.
And if your two reports involve trackers or bifacial modules, there is a published baseline — it just has to be cited carefully. The PVPMC blind intercomparison analysed by Theristis and colleagues simulated six systems across two sites, Albuquerque (New Mexico) and Roskilde (Denmark), and the scenarios were chosen deliberately “to include fixed and tracking systems, monofacial and bifacial modules, modules representative of the current PV market and upcoming technologies, and distinctively different geographical locations/climates”. What does not exist is a published tolerance per configuration: nobody has published how far apart two reports should be for a plant with trackers, bifacial modules and sloping ground at once. IEA PVPS Task 13 ran a dedicated bifacial-tracker exercise over seven systems in 2023, but its combined results are not published in a form we can cite (exercise materials, PVPMC). That distinction — aggregate evidence exists, a per-configuration tolerance does not — is what is worth saying in a meeting, and it is more defensible than “there is no data”.
Layer 2 — the assessor is a bigger variable than the tool
This is the finding that changes how you run the conversation. IEA-PVPS Task 13 ran a blind exercise in which five to seven professional yield assessors received the same inputs for the same plant:
| Site | P50 range (kWh/kWp) | P90 range | σ between assessors | Measured |
|---|---|---|---|---|
| Bolzano, Italy | 1,095 – 1,405 | 997 – 1,274 | 9.7 % | 1,275 (2010–2019) |
| Alice Springs, Australia | 1,757 – 1,985 | 1,631 – 1,819 | 3.9 % | 2,075 initial; 1,926 at 10 years |
And the mechanism, which is the part to remember: the GHI-to-plane-of-array transposition alone ranged from +18.8 % to +24.6 % — “even if the yield assessors all claimed to have used the Perez model” (IEA-PVPS T13-18:2020).
Same model name, same inputs, six points of spread in the first step of the chain. On complex terrain the dispersion between people (≈9.7 %) exceeded the total uncertainty budget any one of them declared (≈6.5 %).
The independent blind intercomparison reaches the same verdict from another angle: across six systems at two sites, the overall median bias was +0.6 % in annual irradiation and −3.3 % in annual energy yield, and while modelling has improved since the 2010 round, “human errors, modeling skills, and derates were found to still cause significant errors” (Theristis et al., Progress in Photovoltaics 31(11), 1144–1157, 2023; abstract).
Practical consequence: the first question is never “which engine is right”. It is “what did each assessor decide”.
And that logic is worth following to the end, even where it does us no favours: if the assessor matters more than the tool, switching tools does not close the discrepancy. What closes it is putting the decisions where they can be argued, one by one. That is why the procedure below asks about decisions rather than brands — and why none of its seven checks names a vendor.
The procedure: seven checks, in this order
Do not start with losses. Start with whether the two numbers even answer the same question.
1. The energy boundary and the period
Identify the delivery point — DC array energy, inverter output, or energy at the point of interconnection — and the period. A specific historical year against a typical meteorological year is a different question, not a discrepancy. Record units, reference capacity (DC or AC) and whether the figure is a base case or a percentile.
2. The weather file, before anything else
Source, dataset version, number of years, gap-filling treatment, time zone and timestamp convention (interval start, centre or end). A half-hour shift in the timestamp convention shows up as a tracker-orientation error later. And note the reference-year trap: TMY files exclude worst-case years, which is why NREL states they “may be more appropriate for preliminary analysis of a system design than for financial decisions” — in their three cases the TMY run came out above the P50 (Dobos, Gilman & Kasberg, NREL/CP-6A20-54488, 2012).
3. Transposition
Given Layer 2, check this before touching electrical losses. Which transposition model, and — because “we both used Perez” proved insufficient — which diffuse decomposition, which albedo value and which horizon treatment.
4. Geometry and the real plant
Pitch, GCR, chord, number of rows per table, tracker limits and backtracking strategy, terrain model. Confirm the compared cases have the same module count and the same DC and AC capacities. A 1 % difference in module count is a 1 % difference in energy that has nothing to do with physics.
5. Stage by stage, not annual totals
Compare equivalent stages: plane-of-array irradiation → array DC energy → inverter output → grid delivery. If the difference already exists at irradiation, stop there. Do not add loss percentages: their reference bases differ, and the same named row can mean different quantities in different tools.
6. Time resolution
Hourly data understates clipping. Measured: using hourly instead of one-minute irradiance underestimates clipping loss by 0.4 % at ILR 1.25 and 1.5 % at ILR 1.5; for a system at ILR 1.23 in San Diego, estimated clipped energy went from below 1 % to above 2.5 % when moving from hourly to one-second data; in Freiburg, inverter efficiency including clipping dropped from 95 % to 94 % going hourly → one-minute, with a further 0.5 % at ten seconds (Anderson & Perry, NREL/CP-5K00-76021, 2020, collecting Good & Johnson 2016 and Burger & Rüther 2006; full text read). If one report is hourly and the other sub-hourly, part of your discrepancy is arithmetic, not disagreement.
7. Change one identifiable assumption at a time
Keep the original case. Investigate one difference, rerun, record what moved. If an effect cannot be represented in both tools, disclose a partial comparison rather than silently disabling it. Four columns are enough: discrepancy · cause investigated · check performed · open question.
Where the physics is genuinely weak — and where the meeting usually isn’t looking
Discussions tend to converge on soiling percentages and availability. The published residuals point elsewhere.
Rear-side irradiance on bifacial plants. Validated against an NREL dataset with front and rear reference cells, relative RMSE was 7.2 % (view factor) and 7.6 % (ray tracing) on the front face, rising to 20.9 % and 22.9 % on the rear. Bias differed in kind: ray tracing −0.1 %, while the view-factor model underestimated the rear by 8.4 % (Ernst et al., Applied Energy 369, 123574, 2024; abstract). One site, one albedo, one racking type — but the order of magnitude is the message: the rear face is modelled roughly three times worse than the front, whichever method you pick.
Module temperature on glass-glass bifacial modules. Two 2026 studies converge: for glass-glass bifacial, the measured heat-transfer coefficient sits near 40 W/m²K, not the 29 W/m²K free-mounted default. In Brazil the best fit was PVsyst with Uc = 40 (RMSE 2.30 K); in Germany, Faiman (RMSE 2.06 K) (Pfeiffer et al., Solar Energy Advances 6, 100140; Kirsten Vidal de Oliveira et al., Progress in Photovoltaics 34(1), 84–103; both abstracts). Both were small fixed systems, not utility-scale trackers, and both conclude that site-specific calibration matters more than the region — so 40 is not a new universal default either.
Backtracking on undulating terrain. Across roughly 7,000 single-axis trackers in seven utility-scale plants, about 1 GW, the controller overrides the ideal backtracking angle to avoid row shading on the undulations, costing ≈2 % of annual energy (0.8–2 % of irradiance, up to 3 % in one case) — and, critically, the loss hides inside the Performance Ratio, which is why it went unnoticed (Santamaría-Sancho et al., Renewable Energy 267, 125734, 2026; abstract). If you have seen the figure “5 %” quoted for this effect, that came from press coverage and includes other operational factors.
And tracker availability, which almost nobody measures. The industry contracts inverter availability — “generally around 99 %” — and tracker availability was left out: historically “trackers have not been included in contractual KPIs as they were considered optimization components”. A 2026 paper proposes the missing KPI and applies it to 64 plants, about 2.1 GWp DC, 80 % of them in European countries, with records from 6 months to 4.9 years. Median availability comes out between 66 % and 88 % depending on how the data is filtered: a median of 66 % when the backtracking period is included, 83 % when only core tracking is evaluated, with the worst plant at 22 % and only two plants above 99 %. The authors set that against what is advertised — “industry claims sometimes even as high as 99 % availability, while the assessment of more than 2 GWp PV power plants tells us a different story” — and call their own results “substantial, and alarming” (Martins et al., EPJ Photovoltaics 17, 3, 2026; open access, full text read).
Two caveats the paper puts up front, and that have to travel with the citation: missing data is counted as unavailability — deliberately conservative, with a median 11 % of data missing in the wider case — and the reference angle is modelled with pvlib’s standard backtracking, which does not reproduce a manufacturer’s proprietary logic, so part of the gap in the wider case “can be partially attributed to modeling errors stemming from the lack of transparency in the control logic”. Even with both concessions, the direction matters for your report: the availability a yield study assumes sits above the availability measured in operation, and that difference lives in none of the four layers above — it lives in the “availability” line almost nobody argues about.
And the finding that reframes all of it: validating the open-source pipeline against well-characterised data, transposition models landed within ±3 % mean bias while thermal models mostly underestimated temperature, with median residuals from −6.5 °C to +2.7 °C, and every thermal model improved when a transient rather than steady-state assumption was used. The authors’ conclusion: “system-specific data availability is more important than model selection”, and using spec-sheet or PAN-file data with complex models “does not guarantee better accuracy than simpler models using system-specific data” (Deville et al., Progress in Photovoltaics 32(5), 291–303, 2024; abstract).
An honesty note about this list. The four items above are also the four things we model most carefully, which is why we know them. The residuals are not ours — they are measured and published by other people, with DOIs — but presenting the list as a neutral survey of the field would be cheating: it is where we have looked. If your discrepancy lives somewhere else, the procedure above works exactly the same.
Read that as the answer to “which tool is more accurate”: on this evidence, the data you feed it beats the model you choose.
Reconciling is not validating
Two simulations that agree can share a bias. Measurements are the other reference, and their quality needs its own review — Sandia’s validation procedure is built around comparing against field data across conditions, not against an annual aggregate (PVPMC).
The fleet check is the humbling one. Plants built after 2015 have generated 7–13 % less than their original P50 estimate, with larger plants (>50 MW) doing worse relative to theirs, and P99 outcomes occurring five times more often than expected with the P50 assumed correct — 17 times more often once P50 optimism is included (kWh Analytics, Solar Risk Assessment; figures via published summaries, we were unable to read the primary PDFs). That gap mixes model error, degradation, availability and weather: it is not evidence about any software product, and anyone who cites it as such, including a vendor, is overreaching.
There is no universal acceptable discrepancy. Agree the scope of the review and publish the differences you could not explain instead of tuning assumptions until the totals match. A reconciliation whose conclusion is “4 % remains unexplained, here is where it lives” is worth more to an IE than one that lands on zero.
Applying this to Heliosolve and PVsyst
A parallel evaluation means rebuilding the inputs in each application and reviewing them; there is no automatic PVsyst project import, and importing a KML/KMZ site boundary is not importing a project. Where the two tools cannot represent the same effect, we would rather show the partial comparison.
What we can put on the table: our documented validation against measured data covers a single plant, so it does not validate every model or location; we have no lender-acceptance track record; and energy and financial results are estimates under your assumptions. See capabilities, limits and working with PVsyst before committing to an exchange, and what our validation does and does not show.
Talk to an engineer and bring one discrepancy you already have. That is a more useful first task than a brand comparison.
Frequently asked questions
How much should two PV yield estimates differ?
There is no standard tolerance. For the simplest configuration — fixed tilt, monofacial, weather data given — seven commercial tools showed mean annual-yield deviations of 2.5 % at laboratory scale and 6.0 % at utility scale. The PVPMC blind intercomparison did include tracking and bifacial modules, across six systems and two climates, with an overall median bias of +0.6 % in annual irradiation and −3.3 % in annual energy yield. What does not exist is a published tolerance per configuration: for a plant with trackers, bifacial modules and terrain at once there is no number you can cite.
Is PVsyst or SAM more accurate?
The published blind comparisons do not rank tools that way, and the open-source validation work concludes that system-specific data availability matters more than model selection. Treat “which tool is right” as the wrong question and check the assessor’s decisions instead.
Why do reports using the same model disagree?
Because model names hide decisions. In a blind exercise where every assessor claimed to use the Perez transposition model, the GHI-to-plane-of-array step still ranged from +18.8 % to +24.6 %.
Does hourly weather data underestimate clipping?
Yes, measurably: about 0.4 % at ILR 1.25 and 1.5 % at ILR 1.5 versus one-minute data, and more at higher resolutions. If one report is hourly and the other sub-hourly, part of the gap is resolution.
Where are yield models weakest today?
On the rear face of bifacial arrays (relative RMSE ≈21–23 % against ≈7 % on the front), in module temperature for glass-glass modules (measured Uc near 40 W/m²K versus the 29 default), and in tracker backtracking over undulating terrain (≈2 %/yr, hidden inside the Performance Ratio).
What tracker availability should a yield report assume?
Not the 99 % that gets advertised. Across 64 plants and about 2.1 GWp, measured median availability landed between 66 % and 88 % depending on data filtering, and only two plants exceeded 99 %. It is a new KPI and its calculation carries caveats — missing data counts as unavailability, and the reference angle is modelled with standard backtracking — but the direction is clear: the availability assumed in a report sits above the availability measured in operation.
Sources. Deville et al., Solar Energy 304:114207 (2026) · Theristis et al., Prog. Photovolt. 31(11) (2023) · IEA-PVPS T13-18:2020 · Deville et al., Prog. Photovolt. 32(5) (2024) · Anderson & Perry, NREL/CP-5K00-76021 (2020) · Ernst et al., Applied Energy 369:123574 (2024) · Pfeiffer et al., Solar Energy Advances 6:100140 (2026) · Kirsten Vidal de Oliveira et al., Prog. Photovolt. 34(1) (2026) · Santamaría-Sancho et al., Renewable Energy 267:125734 (2026) · Martins et al., EPJ Photovoltaics 17:3 (2026), open access · IEA PVPS Task 13 bifacial-tracker modeling exercise (2023), materials at PVPMC · Sandia PVPMC validation procedure · kWh Analytics Solar Risk Assessment (via published summaries). Where we could only read an abstract or a summary, the text says so.
PVsyst, RatedPower, PVcase, SolarFarmer, Solargis, PlantPredict, 3E SynaptiQ and SAM are trademarks of their respective owners. Heliosolve is not affiliated with, endorsed by or sponsored by any of them. Product names are used only to identify the products discussed.