Solar Simulation Accuracy: Software vs Real Output

Last updated: November 2026

SurgePV publishes this post and sells solar design software.

Solar simulation accuracy decides whether a project pays back on schedule or quietly bleeds revenue for two decades. Take an illustrative case: a 7,500 kWp commercial array forecast at 11.2 GWh per year that actually produces 10.2 GWh. That 8.9% gap is worth about $90,000 a year at an assumed $90 per MWh, and the loss repeats every year of the asset’s life. This guide looks at what independent evidence says about PVsyst, HelioScope, SAM and other simulators, why forecasts and field results diverge, and what installers and developers can do to tighten the loop between the two.

What this guide covers:

• What independent sources say about the accuracy of PVsyst, HelioScope, SAM and PV*SOL
• The kWh Analytics underperformance numbers and what drives them
• How P50, P90 and P99 quantify forecast uncertainty
• Eight reasons simulations diverge from real production
• IEC 61724-1 performance ratio analysis and how to close the gap on your own projects

2026 State of Solar Simulation Accuracy

The headline finding from the 2025 kWh Analytics Solar Risk Assessment, as reported by PV Tech in June 2025, is that US solar assets underperform industry-wide P50 estimates by 8.6% on average across more than 34,000 system-months of data from 2015 to 2023. An earlier kWh Analytics study of 2016 to 2019 operating years, reported by PV Tech in October 2020, found an average of 6.3% and said about a quarter of projects missed targets by more than 10% even after adjusting for weather.

This is not mainly a story about bad software. The 2020 report attributes the problem to market growth, new technologies, more complex site designs and a competitive landscape that rewards optimistic forecasts. It also notes that without accepted standards, forecasts vary depending on who runs the model.

Where Real Output Lands Versus Forecast

What the Reports Say About Causes

• Overly optimistic forecasts. The 2020 coverage describes forecasts that became optimistic as the market matured.
• Equipment and operations. The 2025 assessment points to module quality and inverter underperformance, per PV Tech’s summary.
• Data quality for AI tools. PV Tech reports that untrained AI models misclassified up to 20% of operational issues, and 40% to 50% of critical categories such as extreme weather and damage losses.
• Climate. Clean Power Research, writing about the same report, estimated a 4.9% power loss over 30 years at an example site in Europe when climate-model data replaced historical typical-year data. It is one site, so it is a warning about method, not a global forecast.

All of these point to the same conclusion: the solar design tools do their job, but the inputs and the operating environment can drift from the assumptions older models were calibrated against.

How Solar Simulation Software Actually Works

Every modern simulation tool, regardless of vendor, runs the same fundamental calculation chain. Understanding this chain is the foundation of judging accuracy.

The Calculation Stack

1. Meteorological input layer. The simulator pulls a typical meteorological year (TMY) or time-series irradiance dataset from a provider such as Meteonorm, SolarAnywhere, PVGIS, NASA POWER, or Solargis. The dataset provides hourly or sub-hourly global horizontal irradiance (GHI), direct normal irradiance (DNI), diffuse horizontal irradiance (DHI), ambient temperature, wind speed, and humidity.

2. Transposition to plane of array. Algorithms such as Hay-Davies, Perez, or Reindl convert GHI to plane-of-array irradiance based on module tilt, azimuth, and surrounding geometry. Different transposition models can shift annual results by a small but noticeable amount.

3. 3D shading model. A 3D scene with trees, buildings, terrain, and adjacent module rows calculates near and far shading at sub-hourly intervals. Tools differ in how they represent geometry and bypass-diode behavior, so shading is one of the places results diverge.

4. Module electrical model. An equivalent circuit converts cell-level POA irradiance and temperature into DC power. Datasheet specifications (Voc, Isc, Vmp, Imp, temperature coefficients) drive the curve.

5. DC system losses. Soiling, mismatch, ohmic wiring, module quality, IAM (incidence angle modifier), and shading aggregate before inverter input.

6. Inverter model. Efficiency curves from California Energy Commission test data or manufacturer specifications convert DC to AC at each timestep. Maximum power point tracking (MPPT) behavior, clipping at the AC nameplate, and standby losses are modeled.

7. AC system losses. Transformer, line, and parasitic losses reduce the energy reaching the meter.

8. Statistical output layer. Annual energy yield in kWh/year, specific yield in kWh/kWp, performance ratio, and optionally P50/P90/P99 confidence intervals based on uncertainty propagation.

Where Real Tools Diverge

The biggest divergences between tools are usually in the shading step and in the default loss values, not in the weather input or the statistical output. Different shading engines and different defaults can shift the same nominal project. That is why same-input cross-checks between tools can agree closely while real-world comparisons across users show wider spreads driven by input variation.

Validation Statistics: How Accuracy Is Measured

Before comparing tool A against tool B, the industry uses a standard vocabulary for validation error. These are the metrics that show up in independent engineer reviews and research papers.

The Five Core Validation Statistics

For example, a tool with a small annual bias but a larger monthly RMSD would be “accurate on annual aggregates but less accurate on monthly resolution.” That distinction matters because debt service is monthly, while yield warranties are usually annual.

Why Monthly Accuracy Is Worse Than Annual

Monthly errors are usually larger than annual errors because positive and negative monthly deviations partially cancel when aggregated to a year. A site that overpredicts in February and underpredicts in August can still land close on the annual total.

This has practical implications. A 12-month operational data sample is the minimum needed to test a model’s bias. Anything shorter is mostly noise.

What Independent Sources Say About Each Tool

PVsyst

PVsyst is the long-standing reference tool for utility-scale finance, and lender independent engineers commonly use it for P50 and P90 yield reports. Its own validation documentation reports annual accuracy within roughly 5% for seven Swiss installations in the 1996 version 2 release, and within 1 to 2% mean bias error over one year for a 10 kWc amorphous system in 2008. It also states that meteorological variability is about 5% and that meteorological inputs are often known only to 5 to 10%, so input parameters are the main source of uncertainty.

Two recurring sources of PVsyst error follow from that:

1. Meteorological data uncertainty. PVsyst can be no more accurate than its weather input.
2. Default loss assumptions. Defaults are starting points. Real soiling, mismatch and availability depend on the site, so calibrate them against measured data.

The documentation cited above refers to validation done during PVsyst’s development. Treat any single published accuracy figure as one site and one dataset, not a guarantee.

The Bankability Argument

Lenders still ask for PVsyst largely for institutional reasons. It has decades of independent engineer review behind it and a stable methodology. For a 25-year asset that will be refinanced, that stability matters. That is why many EPCs run PVsyst vs HelioScope in a complementary workflow: HelioScope for fast layout iteration, PVsyst for the final investment-grade yield report.

HelioScope

HelioScope is a web-based simulation tool. In 2014, PV Tech reported that DNV GL analysed HelioScope against PVsyst as the benchmark and found agreement with PVsyst results to within 1%. That is agreement between two models, not between model and measured output, and PV Tech notes DNV GL’s wording on shading was tentative. The same article cites a separate PsomasFMG evaluation, comparing predictions with real measured output, that put the average difference at 0.26%. That figure is a company-reported result, so treat it as indicative.

NREL has also published a technical report that tests HelioScope’s simulation engine against measured PV system data and compares it with other tools (NREL/TP-6A20-72155). Read the report for the results rather than relying on a summary.

For deeper detail on HelioScope’s strengths and constraints, see our HelioScope review. Confirm current project-size limits and bankability status with the vendor, as those change over time.

SAM, Aurora Solar and PV*SOL

NREL’s System Advisor Model is free and open source, and it is widely used in research and government project evaluations. Its workflow suits engineers more than a sales pipeline, so it is rarely the production tool for an installer.

Aurora Solar is a widely used design platform for residential work, and the Aurora Solar vs HelioScope comparison covers how the two differ in workflow. We did not find a public independent benchmark of Aurora’s yield output against PVsyst at C&I scale, so ask the vendor for validation documentation if you need it.

PV*SOL from Valentin Software is widely used in Europe. We did not find independent validation data for it that we could verify, so check Valentin’s own documentation and any independent engineer’s report for your market.

Side-by-Side: What Each Tool Publishes

P50, P90, P99: How Software Quantifies Uncertainty

Predicting a single annual yield number is only half of what simulation software should do. The other half is quantifying how confident the forecast is. That is where P-values come in.

Definitions

• P50: Median forecast. The plant has a 50 percent probability of producing more than this value in any given year. This is the central estimate.
• P75: Conservative central case. Exceeded in 75 percent of years.
• P90: Downside stress test. Exceeded in 90 percent of years. Lenders size debt against P90.
• P95: Deep stress case. Exceeded in 95 percent of years.
• P99: Worst-case bound. Exceeded in 99 percent of years.

PV Magazine’s June 2025 explainer describes the same idea: P50 is the central estimate, and a P90 means a 90% likelihood that energy will meet or exceed the stated number, with 10% below it.

The mathematical relationship is straightforward when uncertainty is approximately normal: P90 equals P50 times (1 minus 1.282 times sigma_total), where sigma_total is the combined uncertainty as a fraction of P50.

Sources of Uncertainty

Combined uncertainty comes from multiple independent contributors:

The table below is an illustrative uncertainty budget. The values are example inputs to show the method, not published benchmarks. Replace them with values from your own data provider and independent engineer.

Combined via root-sum-square (assuming independence), these example values give sigma_total of about 6.4%: the square root of 9 + 16 + 6.25 + 4 + 4 + 1.5625. That puts P90 about 8% below P50 (1.282 × 6.4%).

Why Uncertainty Is Often Underestimated

Uncertainty is often underestimated. In its 2023 analysis reported by POWER Engineering, kWh Analytics found that annual performance relative to P50 is strongly left-skewed compared with the normal distribution typically assumed, so P99 outcomes, nominally 1-in-100-year events, were showing up about every 20 years. A normal-distribution P90 can therefore be too optimistic.

This matters because P90 is the number lenders use to size debt. Underestimating P90 means projects are over-leveraged relative to their true downside risk.

For a deeper treatment of the math, the P50 P90 energy yield explained guide breaks down the calculation step by step. The P50/P90 glossary entry provides a one-page reference.

The Two-Stage Uncertainty Decomposition

NREL recommends decomposing total uncertainty into two stages:

1. Interannual weather variability. Cannot be reduced by better modeling, only by longer site data records.
2. Model and parameter uncertainty. Can be reduced by better meteorological inputs, better shading scenes, and tighter loss assumptions.

If model uncertainty dominates weather variability for a given project, investment in better tooling pays off. If weather variability dominates, the project needs more data history, not better software.

The 8 Most Common Sources of Prediction Error

When a project underperforms, the source is rarely the simulation engine itself. It is usually one of these inputs. The sizes of the effects depend on the site, so measure them rather than trusting any generic range.

1. Meteorological Data Quality

The meteo dataset is the foundation. A TMY file is a synthetic year that aims to represent average long-term conditions, and PVsyst’s documentation notes that meteorological inputs are often known only to 5 to 10%. Switching from a free dataset to a ground-calibrated one can shift modeled yield noticeably, so test both on a project where you have measured data.

2. Shading Scene Fidelity

Missing trees, water towers, parapets, and HVAC equipment is a common error in C&I models. A model with default flat-terrain assumptions on a site with a tall parapet on the south edge can overstate first-row yield.

For complex sites, professional solar shadow analysis software with accurate 3D geometry removes much of the input error. SurgePV’s shading analysis works on the same design model as the layout, so you can compare options before you finalize.

3. Default Soiling and Snow Losses

Every tool ships default soiling values. Real-world soiling ranges from negligible in rainy, clean climates to large in dusty desert and agricultural sites, and snow losses matter at northern latitudes. Calibrating these against measured field data is essential. The soiling loss glossary covers the math and field measurement methods.

4. Module Quality and Mismatch

Default module mismatch losses assume tight binning. For modules with wide bin tolerance or mixed shipments, real mismatch can be higher. The 2025 kWh Analytics findings on module quality are partly explained by this.

5. Inverter Availability

Default inverter availability in many tools is 100 percent (zero downtime). Real availability is lower and depends on inverter generation, climate, and maintenance. PV Tech’s summary of the 2025 assessment points to inverter underperformance as a contributor, so consider loosening the default for forecast purposes.

6. Degradation Modeling

A default degradation rate may understate performance loss for some module shipments. Over a 25-year asset life, a small difference in annual rate compounds into a meaningful gap by year 20. The solar panel degradation rates guide breaks this down by technology.

7. AC Side Losses

Transformer losses, line losses, and parasitic losses are often left at defaults. For long DC runs, undersized cables, and oversized transformers, real AC losses can be higher. The solar system losses guide covers the full loss chain, from STC nameplate to AC meter.

8. Operational Reality vs Simulation Assumptions

Curtailment, reactive power requirements, frequency response obligations, and grid feeder voltage variation are rarely modeled in default simulations. On constrained feeders, curtailment alone can account for a noticeable annual yield loss. Note that the 2020 kWh Analytics figure of 6.3% excluded curtailment.

Performance Ratio: The Field Metric Behind Validation (IEC 61724)

Once a project is operating, the question shifts from forecast accuracy to verification. The performance ratio (PR) defined in IEC 61724-1 is the industry-standard metric for comparing actual production against expected output.

The IEC 61724-1 Performance Ratio Formula

In its simplest form:

PR = (E_AC / P_nameplate) / (H_POA / G_STC)

Where:

• E_AC is the measured AC energy output in kWh over the analysis period
• P_nameplate is the DC nameplate rating in kW at STC
• H_POA is the measured plane-of-array irradiation in kWh/m² over the period
• G_STC is the standard test condition irradiance, 1 kW/m²

The PR is dimensionless and typically expressed as a percentage. A PR of 85 percent means the system delivered 85 percent of the energy it would have produced if every module operated at its STC nameplate rating across the measured irradiance.

What Drives Performance Ratio

Typical performance ratio depends on climate, system type, design and monitoring class, so compare your measured PR with your own simulated PR rather than a generic benchmark. A residential rooftop with heavy shading, a commercial rooftop, a ground-mount plant and a tracker plant will all land in different places. The table below shows what usually drives the difference.

Weather-Corrected PR

The base PR formula bakes weather variability into the ratio. To compare a single plant across seasons or against another plant in a different climate, the weather-corrected PR adjusts the denominator for actual module temperature, using the module Pmax temperature coefficient (typically negative 0.0035 to negative 0.0045 per degree Celsius for crystalline silicon).

This is the form recommended by NREL and described in the NREL work on weather-corrected performance ratio, and IEC 61724-1 covers the performance ratio and its monitoring requirements. For detailed methodology, the solar performance ratio guide walks through both the standard and weather-corrected forms with worked examples. The how to calculate solar performance ratio post covers data collection requirements and common measurement pitfalls.

Why PR Matters for Closing the Forecast Loop

The simulated PR appears in every PVsyst, HelioScope, and SAM yield report as a single headline number. After 12 months of operation, the measured PR can be calculated using IEC 61724-1 methodology and compared directly to the simulated value. The gap is the model error.

For a 100 kWp commercial system simulated at PR 84 percent that delivers measured PR 78 percent across the first operating year, the 6-percentage-point gap is the calibration error that should flow back into the next project’s loss assumptions. This is how disciplined EPC organizations close the loop between forecast and reality.

IEC 61724-1 defines monitoring classes A, B and C with different sensor and accuracy requirements. The class is set in the engineering procurement contract and flows into the asset valuation, so choose it early. Underestimating monitoring requirements in design often shows up later as an inability to validate warranty claims.

Improving Your Own Forecasts: A Practical Workflow

For installers and developers building their own simulation discipline, five practices have outsized impact on prediction quality.

1. Upgrade Meteorological Inputs

Move from free datasets to a paid or ground-calibrated dataset for larger projects, and for utility-scale work invest in on-site irradiance instrumentation during development. Test the change on a project with measured data to see whether the accuracy gain covers the subscription.

2. Build Full 3D Shading Scenes

Model trees, terrain, adjacent buildings, parapets, and HVAC obstructions. For complex roofs, integrate LiDAR or drone-captured 3D scans. Default flat-terrain assumptions are a common source of optimistic yield error in C&I models. SurgePV’s solar design software models 3D roofs and shading in the design canvas so you can compare layout options.

3. Calibrate Loss Assumptions Against Measured Data

Maintain a library of measured PR data from your operating portfolio. After each project’s first 12 operating months, compare measured PR to simulated PR and update the soiling, snow, availability, and degradation defaults for the next project in the pipeline. This practice can close more of the forecast gap than a software upgrade.

4. Cross-Validate Across Two Independent Tools

For high-stakes projects, run the design in both HelioScope and PVsyst. If the two agree within a couple of percent on annual yield, the model is internally consistent. If they disagree by more, the cause is almost always an input difference: meteo, shading, or loss assumptions.

Here is a hypothetical example. Suppose HelioScope returns 3,432 MWh/year and PVsyst returns 3,397 MWh/year for the same 2.1 MWp rooftop with the same weather file, equipment and losses. The gap is 1.0%, which is within the cross-tool noise you would expect, and might trace to the transposition model. If you then switch the weather source in only one tool and the gap jumps to about 4.7%, the input layer, not the physics, is driving the divergence. The numbers are illustrative, not from a real project.

5. Publish P50 and P90 Together, Not Just P50

For any project headed to financing, present P50 and P90 yield estimates together in the proposal. Lenders will calculate P90 themselves if the developer does not, and the lender’s P90 will be more conservative than the developer’s. Presenting both upfront builds credibility and accelerates the financing decision.

Workflow Checklist

SurgePV supports the design and proposal side of this workflow: 3D roof modeling, array layout, shading analysis, energy-yield and financial modeling and proposal generation. It does not replace an independent engineer’s bankability report.

Closing the Forecast Gap: Action Items

Solar simulation is more accurate than it has ever been, yet the field gap between forecast and reality persists because input discipline has not kept pace with software capability. The tools are not the problem. The inputs are.

• Audit your meteo data sources. If you default to a free dataset for larger projects, test a ground-calibrated one on a site where you have measured data.
• Build a measured PR feedback loop. After every project completes its first operating year, calculate IEC 61724-1 PR and update your loss-assumption library. This is the highest-impact practice for closing the forecast-to-reality gap.
• Stop quoting P50 alone. Always present P50 and P90 together. If your software does not output P90 natively, run the uncertainty math separately.

For installers benchmarking their own toolchain, the best solar software for simulation roundup is a starting point. Choosing the right solar software is less about feature lists and more about how well the toolchain handles the input discipline that determines whether forecasts hold up in operation.

Frequently Asked Questions

How accurate are solar simulation software predictions compared to actual production?

When inputs are well calibrated, annual predictions from established tools can land close to measured output, but many real projects still miss their forecasts. The 2025 kWh Analytics Solar Risk Assessment found US solar assets averaged 8.6% below industry-wide P50 forecasts across more than 34,000 system-months. The gap is mostly an input and assumption problem, not a core software error.

Is PVsyst more accurate than HelioScope?

On the same inputs, DNV GL found HelioScope’s results agreed with PVsyst’s within 1%, per a 2014 PV Tech report. That is agreement between two models, not proof of accuracy against measured output. The practical difference is workflow: PVsyst is the long-standing reference for lender reports, while HelioScope is built for faster commercial design iteration.

What is the average gap between simulated and real solar output?

The 2025 kWh Analytics Solar Risk Assessment, covering more than 34,000 system-months from 2015 to 2023, found US PV assets 8.6% below industry-wide P50 forecasts on average. An earlier kWh Analytics study of 2016 to 2019 operating years found 6.3%. The samples differ, so treat the two figures as context rather than a clean trend.

How do P50 and P90 affect solar project financing?

P50 is the median yield forecast, equally likely to be exceeded or missed in a given year. P90 is the conservative case, exceeded 90% of the time. Lenders size debt against P90 because it represents the financial stress case, so understating uncertainty can leave a project over-leveraged.

How can installers improve solar simulation accuracy?

Five practices matter most. Use site-specific meteorological data rather than nearest-station defaults. Build a full 3D shading scene. Calibrate soiling, snow and degradation losses against measured data from comparable installations. Cross-check results in a second tool. After commissioning, compare measured performance ratio under IEC 61724-1 with the simulated value and update your assumptions.

Why does real solar output fall short of simulation?

Most shortfalls come from optimistic inputs rather than the simulation engine: soiling, snow, inverter availability, degradation, curtailment and shading scene errors. kWh Analytics also reports module quality and inverter problems as contributors. Calibrating the loss tree against measured data closes much of the gap.

To see how layout, shading and yield modeling fit together in one project, book a SurgePV demo.