PV degradation uncertainty and lifetime#

Implements the methodology in Kaaya, Ascencio-Vásquez, Weiss and Topič (2021), Assessment of uncertainties and variations in PV modules degradation rates and lifetime predictions using physical models, Solar Energy 218, 354–367, DOI: 10.1016/j.solener.2021.01.071. The supplied paper is the source; its original field time series are unavailable.

The implementation separates reusable methods from an offline analysis and executable gallery example. It does not replace the existing 2019 climate degradation mapping.

Method and input contracts#

Module

Paper coverage

Inputs and outputs

energy_methods.pv_reliability

Table 2, Eqs. 19–20

Four power-versus-age curves, analytic threshold crossing, curve fitting, NRMSE and symmetric relative difference

energy_methods.degradation_uncertainty

Eqs. 1, 3–6

Daily stress-product rate and combined hydrolysis/photo/thermal rates, explicit native-to-fractional/year scale

energy_methods.pv_stress_weather

Eqs. 7–17

Ross/Faiman temperature and fitting; linear, original/modified Wald and air-mass polynomial UV; Buck RH

energy_methods.degradation_preprocessing

Sections 2.5 and 3.4

Power screening/normalization and complete-year hourly weather to annual stresses

energy_analysis.degradation_uncertainty

Tables 3–6, 8, 10–11

Rounded-table checks, one-at-a-time climate scenarios, synthetic calibration and weather-model comparisons

Models accept prepared inputs; acquisition stays in data-retrieval. Use the existing ERA5/PVGIS connectors as appropriate, retaining native measurement height, spectral band, POA geometry, units, source dates and quality flags. Do not substitute ambient temperature for module temperature, GHI for POA, or annual UV dose for daily irradiance. ERA5 UV through 440 nm is not the study’s 280–400 nm UV band. No new remote connector is needed for this implementation.

Reliability and coefficient units#

With age t in years and power normalized to initial power:

API kind

Normalized power

Time to retained fraction q

linear

1 - k*t

(1-q)/k

exponential

exp(-k*t)

-ln(q)/k

stretched_exponential

exp(-k*t**beta)

(-ln(q)/k)**(1/beta)

wear_out

1-exp(-(gamma/(k*t))**mu)

gamma/(k*(-ln(1-q))**(1/mu))

k is fractional, not a numerical percentage: Table 3’s 0.83 is entered as 0.0083. Stretched-exponential k has units year**(-beta), despite the paper’s common percent/year label. Comparing the four coefficients as if they were identical annual loss rates is misleading. Report the complete curve and shape. Threshold crossing is a deterministic extrapolation, not a survival distribution, confidence interval, field-validated service life or warranty.

At age zero, all curves return one. Zero rate gives infinite threshold time. Linear curves are not clipped after zero power. fit_reliability fits k, and beta or mu where appropriate, using deterministic multistart bounded least squares. Wear-out gamma is fixed by the caller: only gamma/k is identifiable, so fitting both would produce an underdetermined model. The returned Jacobian condition number exposes poor conditioning; it is not parameter uncertainty. Shapes are bounded to 0.05–10 and k to 1e-10–10 as numerical fitting choices.

from energy_methods.pv_reliability import ReliabilityCurve, fit_reliability

curve = ReliabilityCurve("stretched_exponential", k=0.003, beta=1.59)
print(curve.failure_time(retained_power=0.8))
print(curve.power([0, 5, 10, 20]))
# Supply measured ages and normalized power, after independent quality screening:
# fit = fit_reliability(ages, power, kind="wear_out", gamma=0.43)

The rounded published coefficients give approximately 24.1, 26.3, 15.0 and 19.8 years. Small differences from the printed 24.0, 26.2, 15.1 and 19.7 are retained, not tuned away. These are checks of published parameter arithmetic; no original observations have been refitted.

Physical-rate conventions and unresolved source details#

The new combined model preserves the printed thermal factor (273 + upper_c - lower_c)**theta in Eq. 6. This differs from the existing 2019 mapping function’s (upper_c - lower_c)**theta. A temperature difference does not physically require a kelvin offset. The new function deliberately implements the printed empirical form without silently declaring it a typographical error. Both paths share validated Arrhenius/combination algebra; existing API behavior and tests remain unchanged. Arrhenius temperatures use Celsius + 273.15.

The combination is A_N * product(1 + k_i) - 1, with subtraction outside normalization. Mechanism outputs are not additive shares. Table 5 gives nine coefficients but omits A_N and cycling frequency C_N; the paper’s dimensional labels also do not fully settle the calibration’s numerical rate convention. The implementation therefore requires all these choices. No generic module calibration is silently introduced.

rate_scale_to_year_inverse is mandatory: 0.01 means the native number has a percent/year convention, 1 means fractional/year, and a genuine per-second calibration would require its declared seconds/year conversion. This conversion is applied after the combined expression; rescaling components before the multiplicative combination changes the model.

The daily stress-product equation uses daily maximum module temperature, temperature range, daytime UV irradiance in W/m², and mean RH in percent. Its beta0 includes calibration-dependent units. The annual dose in Table 6 cannot replace that irradiance. The example exercises both physical equations with synthetic hourly weather and discloses different aggregation paths and assumed scales. It does not claim a fair ranking or reproduce Figure 3. Physical-rate coefficient refitting across sites is not provided: the original matched performance/weather records and complete calibration conventions are needed for credible calibration and independent-location validation.

Weather, filtering and uncertainty#

Temperature coefficients are explicit; Faiman wind is supplied at the height used in calibration, without a hidden height correction. Calibration uses one year and the synthetic demonstration evaluates on another. UV polynomial coefficients must be supplied; the code implements prediction, not a new UV coefficient optimizer. Modified Wald preserves the a-b*kt sign convention, including negative b values printed in Table 9. Invalid UV fractions fail instead of clipping silently. The Buck prefactors are in kPa, despite the paper’s nomenclature saying Pa; the ratio for RH is unchanged. The accepted Buck domain is -80 to 60 °C; dewpoint above ambient fails.

Power preprocessing keeps 800–1200 W/m² inclusive, scales power to 1000 W/m², and applies a mean ± 2 population-SD filter. It does not correct temperature to 25 °C. Monthly means, centered rolling windows (default 12 months), mid-month age labels and max-trend normalization are documented repository choices where the paper does not specify an executable configuration. Centering uses future data: use for retrospective trend analysis, and prepare calibration and holdout intervals separately. Missing months remain gaps. A UTC age origin is required; removing early observations does not reset module age. Filters cannot establish that soiling, snow, outages or sensor changes have been removed.

Annual processing requires an entire UTC hourly calendar year (including leap day) and finite values. It uses POA>0 daytime temperature/RH means and the 5th/95th temperature percentiles; UV integrates every hourly interval. Cycling frequency is independent of the number of hours. This is different from the existing mapping helper’s averages of daily extrema. Multi-year/TMY assessment and independent module/site validation remain necessary for deployment.

NRMSE follows Eq. 19 using the measured mean, or measured range when the mean is exactly zero. Negative measured means fail; the all-zero reference is defined only for an exact match. Eq. 20 uses 100*abs(p-m)/max(abs(p),abs(m)); a zero/zero pair is defined as zero. These are error/variation metrics, not probabilistic uncertainty intervals. Inputs must align; missing pairs are not silently dropped.

Run the reproducible analysis#

uv sync --locked --all-packages
uv run python -m energy_analysis.degradation_uncertainty \
  --output artifacts/degradation-uncertainty
uv run python examples/analysis/plot_degradation_uncertainty.py

The study writes six CSVs, two PNGs and a manifest with input/output SHA-256, source metadata, numerical assumptions and software versions. Published-table facts, SCENARIO outputs and SYNTHETIC results are identified separately. The JSON table fixture is packaged in the wheel, so both commands run offline outside the checkout after installation. No article PDF or original figures are redistributed.

The default climate scenario declares A_N=1, 365 cycles/year and a 0.01 conversion factor. These are unvalidated example assumptions. Temperature, RH or UV inputs are replaced one at a time; other inputs and curve shape stay fixed. They do not reproduce the paper’s stated 1.7–14.5% variations. The synthetic weather generator uses Faiman/Wald/Buck, so their low errors are true by construction, not field evidence for their superiority.

Source inconsistencies stay visible: Negev’s UV is 96.7 kWh/m² in Table 6 and 96.9 in Table 10. Each comparison uses its own table’s reference. Some printed relative differences also disagree with Eq. 20 on the rounded inputs: for Table 10’s 96.9 versus 96.5, the implementation returns about 0.413%, not the printed 0.2%. The raw time series would be required to investigate further.

Validation covers analytic inverse crossings, identifiable calibration recovery, unit scaling, the two thermal conventions, normalization placement, weather calibration, source-table checks, leap years, missing intervals, trend gaps and byte-identical CSV replay. It verifies software behavior, not field prediction accuracy or all numerical claims in the publication.