The Real Cost of a PV Simulation Error
A 3% bias in modeled energy doesn’t cost 3% of the project. On a 100 MWdc plant it can strip €1.3M of bankable debt and push the debt service coverage ratio (DSCR) from 1.30 into technical default. The error starts in a thermal or shading model; it ends in the capital stack. Here is the mechanism, with the numbers.
TL;DR
- Annual energy sits only in the denominator of LCOE: a 1% energy error is a 1% LCOE error — and up to 1.58% on NPV (Soto Calvo & Lee, 2025).
- Lenders don’t size debt to the mean (P50). They size to a downside case (P90/P99). Overestimating yield oversizes the debt.
- The cost is asymmetric: requiring P90 is equivalent to a lender pricing overestimation 9× worse than underestimation.
- The real fleet already runs closer to P90 than P50: −6.3% on average, and 7–13% below on post-2015 projects (kWh Analytics).
- The lever isn’t a higher P50 — it’s shrinking the model’s systematic bias, the only thing a lender actually pays for.
Developers, EPCs, and independent engineers all live on one number: expected annual energy. The entire financial model hangs from it. Yet that number comes out of a simulation whose error bars are rarely translated into money. Let’s translate them.
Why a simulation error is a financial error
Because energy is the denominator of nearly everything. The levelized cost of electricity (LCOE) is lifetime cost divided by discounted lifetime energy. Energy appears only on the bottom, so the elasticity is exact: −1. A 1% drop in energy is a 1% rise in LCOE (Micheli et al., 2024; Mandys et al., 2023).
On net present value the effect is larger than one-to-one. Because costs are mostly fixed, the full energy error lands on the margin: global sensitivity studies rank capacity factor as the #1 NPV driver, with an elasticity of ≈1.58 — ahead of CAPEX (1.25) and price (1.24) (Soto Calvo & Lee, 2025).
In this article’s base case — 100 MWdc, a €40/MWh PPA — a +3% energy bias shows +5 GWh/year that don’t exist. Via the 1.58 elasticity, that’s on the order of a 4.7% NPV swing on a phantom figure. The technical error is already a valuation error.
Where do the most expensive errors slip in?
Not every error weighs the same: the dangerous ones are those that don’t average out over the years. Interannual weather variability dilutes as the plant accumulates operating years (as 1/√n). Model bias does not — it is a permanent floor. And that’s exactly where the money is.
The industry’s reference uncertainty budget (IEA-PVPS Task 13, 2020) ranks the bias sources like this:
| Bias source | Typical magnitude (1σ) | Averages out over years? |
|---|---|---|
| Resource / transposition to plane-of-array | 2.5–4.0% | No |
| Reflection (IAM) / spectral | 0.5–3.0% | No |
| Shading (near, far, inter-row) | 0.5–3.0% | No |
| Thermal model (cell temperature) | 0.5–2.0% of yield | No |
| Soiling | 0.5–2.0% | Partly |
| Inverter clipping (sub-hourly) | +1–4% bias | No |
| Degradation rate (compounded 25 yr) | ≈0.3%/yr | No (accumulates) |
Validation against plant data returns biases of −1.7% to +5.5% depending on the tool (Freeman et al., NREL 2014) — proof this isn’t theoretical: it’s the gap between commercial software on the same asset. And the human assessor is itself a source of error: seven engineers given identical inputs produced P50 estimates with a 9.7% standard deviation (IEA-PVPS, 2020).
The domino effect on financing
This is where the energy error becomes euros of debt. And the mechanism carries an asymmetry worth understanding before you sign.
What is P90, and why do lenders use it?
P90 is the production expected to be met or exceeded in 90% of years. It’s computed by shifting the P50 (the median) down by the total uncertainty: P90 = P50 · (1 − 1.282 · σ) (Solargis; NREL/SAM). At our case’s 6.5% uncertainty, P90 lands 8.3% below P50.
Lenders do not lend against the mean. They size debt to the lesser of two cases — P50 at a target DSCR and P99 at 1.00 DSCR — and take whichever binds (Pivotal180; Renewables Valuation Institute). Overestimating production inflates both cases, and with them the debt that gets committed.
The 100 MW case, step by step
| Base-case scale | Value |
|---|---|
| Energy P50 | 170,000 MWh (170 GWh) |
| Total uncertainty σ | 6.5% (IEA-PVPS 2020) |
| P90 = 170,000 · (1 − 1.282·0.065) | 155,833 MWh |
| P99 = 170,000 · (1 − 2.326·0.065) | 144,298 MWh |
| Unit margin (PPA 40 − opex 10) | €30/MWh |
| Cash available for debt service (CADS) P50 | €5.10M/yr |
| Target DSCR P50 / P99 | 1.30 / 1.00 |
| Sustainable debt (18-yr tenor, i 5%) | ≈ €45.9M |
Now inject a +3% optimistic bias. The simulator reports a P50 of 175,100 MWh. The lender sizes against that inflated cash flow and lends ≈ €47.2M: €1.3M of debt the asset cannot sustain. When reality delivers the true 170 GWh, the realized DSCR falls from 1.30 to 1.26; and in a downside year (true P99) it drops to ≈1.07 — versus 1.10 for the unbiased design — crossing the cash-sweep threshold (<1.15) and grazing technical default (<1.05).
The marginal cost of erring high is 9× the cost of erring low. That’s not a metaphor: it’s what sizing to P90 means. Formally, requiring P90 is equivalent to a lender pricing overestimation at 9-to-1 versus underestimation (P99 ⇒ 99-to-1). Overestimation triggers covenant breach, liquidated damages, and equity haircuts; underestimation only costs the upside you didn’t lever.
Market data point: under P90 realities, equity cash yields are “cut in half” for the life of the asset (Norton Rose Fulbright, 2020, citing kWh Analytics data).
The cost of finding out late
Finding the error in operating year 3 doesn’t reduce it — it only confirms the debt was mis-sized in year 0. Two multipliers make it worse.
First, compounding. A 0.3%/yr error in the degradation rate (Jordan & Kurtz, 2013) accumulates over 25 years into several percentage points of lifetime energy. Second, irreversibility. The inverter loading ratio (ILR), cable cross-section, and electrical topology are frozen at design. A clipping bias that grows with ILR (Anderson et al., NREL/NextEra, 2022) is locked into steel and copper.
Weather variability does average out over the years; model bias does not. The real fleet proves it: post-2015 projects generate 7–13% below their P50 and run “closer to P90,” with a 3–5% optimistic bias already documented by the independent engineer (DNV; kWh Analytics, via Norton Rose Fulbright).
How to cut the risk at the modeling stage
The lever isn’t in the weather — which time corrects for free — but in the systematic bias. Narrowing the uncertainty band raises the bankable quantile P90 = P50·(1 − 1.282·σ) and, with it, the debt and the equity return.
And the magnitude is measurable. Cutting resource uncertainty from 2.29% to 1.20% adds ≈$175,000/yr of cash available for debt service on a 150 MW plant, and improves the P90 DSCR by +0.02× (SolarAnywhere, 2026). In leverage terms, moving from ±10% to ±8% uncertainty lifts LTV from 70% to 72% (Solargis). Site adaptation with ground measurement — which compresses satellite bias from ±3.5% to ±2–2.5% — is the highest-return lever per euro invested.
No simulation is exact. Being honest about where the error bars live — weather-year selection, soiling, degradation, thermal assumptions — is part of the job. The goal isn’t to eliminate uncertainty; it’s to bound the term that doesn’t dilute with time: model bias.
How Heliosolve solves it
A simulator’s value isn’t measured by the P50 it reports, but by how far it lowers the systematic uncertainty floor. And that floor is the sum-in-quadrature of several biases, so lowering it means attacking all of them — not just the largest. Heliosolve does this across the four heaviest sources in the IEA-PVPS budget:
- Advanced irradiance model → attacks the resource term, the largest in the budget (2.5–4%). It’s the highest-return lever because it dominates the quadrature sum.
- Optical ray tracing → resolves the real geometry of direct, diffuse, and reflected light (including bifacial gain) instead of isotropic transposition models and lumped shading factors, whose model-to-model scatter runs from <5% to >10% of plane-of-array irradiance.
- Cell-level computation → captures the electrical mismatch and the shading–thermal–electrical coupling that lumped loss factors ignore. It’s what keeps the precision of the ray tracing and thermal model from washing out into a single plant factor.
- Heat-transfer thermal engine → solves a wind-sensitive heat balance instead of NOCT-type correlations — the source NREL flags as one of the largest after the resource, and one that moves annual yield by −0.74% to −1.85% per +2 to +5 °C of cell-temperature error.
All four share one goal: drive the bias toward zero at every link in the loss chain. Because, by the P90 math, every point that lowers that floor raises bankable energy — and with it, more debt at a lower cost. Heliosolve doesn’t promise a higher P50; it promises a bias closer to zero, which is the only thing a lender pays for.
Frequently asked questions
What is P90 in a solar production study?
P90 is the annual energy level expected to be met or exceeded in 90% of years. It’s obtained by subtracting 1.282 times the total uncertainty from the P50 (the median): P90 = P50 · (1 − 1.282 · σ). Lenders use it as the base case for sizing debt because it represents a prudent scenario.
Why is overestimating production worse than underestimating it?
Because financing is sized to a downside case. Overestimation oversizes the debt and the commitments (PPA, DSCR), triggering covenant breach, liquidated damages, and equity haircuts. Underestimation only costs unlevered upside. Requiring P90 is equivalent to penalizing overestimation 9× more.
How much does a 3% simulation error cost in money?
On a 100 MWdc plant with a €40/MWh PPA, a 3% optimistic bias can oversize the debt by ≈€1.3M and push the realized DSCR from 1.30 into cash-sweep or technical-default territory in low-production years. The exact figure depends on the structure, but the order of magnitude is six figures.
What uncertainty is normal in a PV yield study?
First-year total uncertainty for a well-characterized plant is around 5–7% (1σ), dominated by the solar resource and transposition. The IEA-PVPS reference budget (2020) gives 6.5% total, of which irradiance contributes ~4% and transposition ~2.5%.
How do you reduce a PV model’s systematic uncertainty?
By attacking the bias, not the weather variability: site adaptation with ground measurement (cuts satellite error from ±3.5% to ±2–2.5%), high-fidelity optical and thermal modeling, and cell-level computation. Only the model term persists over the years; it’s the one where precision pays.