Climate Change Yield Assessment (CCYA)#

CCYA estimates how a PV system’s specific DC yield changes under supplied climate projections, separating irradiance, temperature and power degradation. The implementation is energy_methods.ccya in the reusable methods package. It fits historical monthly observations to a climate model and projects future months for one site, climate model, ensemble member and scenario at a time. It makes no network requests and requires no new dependencies.

Scientific source and implementation boundary#

Primary source: Julián Andrés Ascencio Vásquez (2021), Modelling and worldwide assessment of performance and aging of photovoltaic modules and systems, University of Ljubljana, Chapter 7, sections 7.5.1-7.5.3, printed pp. 150-164. The CCYA definition is on p. 156, Figure 7.10. Its Bolzano application is on pp. 157-164. Thermal expressions appear in equations 3.6 and 3.8, p. 52.

The thesis cites a CCYA manuscript as to be submitted, bibliography [8]. This implementation cites the thesis; it does not invent a published CCYA article or DOI. The Ross temperature relation is also credited through Skoplaki and Palyvos (2009), thesis reference [154]. Polo et al. (2020), thesis reference [287], is credited as background for site adaptation, without claiming to reproduce all techniques benchmarked in that paper. See the reference catalogue and the module’s REFERENCE_IDS.

This is a documented implementation of the CCYA workflow, not a numerical reproduction of the Bolzano results. Original author code, plant measurements, selected climate files and all fitting settings were not supplied. The thesis does not fully specify the seasonal fitting rule, percentile convention, temperature normalization or rolling-window alignment. The choices below are part of this implementation and must be retained when reporting results.

Thesis element

Implementation

Section 7.5.1: linear site adaptation

Ordinary least squares measured = slope * modelled + intercept on the explicitly supplied monthly overlap, separately for GHI and ambient temperature.

Figure 7.13: monthly centers of modelled GHI and measured mean daily DC yield

Twelve through-origin proportional responses, one for each calendar month, passing through that month’s centroid. This fitting rule is an explicit project choice.

Figures 7.14-7.15: annual yield and 10-year mean

Sum native-calendar monthly yields; trailing 10 complete years, including the reported year. Shorter windows stay missing.

Equation 7.1: rolling change relative to measured long-term mean

100 * (rolling annual mean / mean measured annual yield - 1). This generalizes the solar-resource expression to DC yield.

Equations 7.2-7.4: cumulative comparisons

Sum annual projected yield and divide by historical annual baseline * operating-year count. Export both ratio (%) and signed gain/loss (%).

Equations 3.6, 3.8; Figure 7.17

Ross module temperature with GHI, then an incremental multiplier relative to the historical calendar-month temperature. Defaults: gamma = -0.0045 /°C, Ross = 0.03 °C m²/W.

Figure 7.19: power retention

Linear degradation sampled at the start of each operating year; year 1 has age zero by default. Explicit aligned monthly factors support nonlinear paths. Rates are supplied, not inferred.

Input contract#

Pass pandas DataFrames with a monthly PeriodIndex(freq="M"). The index is a year/month label, not a UTC timestamp. Every table must contain consecutive, complete January-December years, in order, with unique months and finite values. Measured and historical-climate tables must cover the exact same selected years. The default minimum calibration is two years; that is an execution guard, not a claim that two years suffice for a reliable climate or financial assessment. Use longer qualified records and chronological holdout evaluation in real studies.

Column

Required in

Definition

ghi_wm2

All tables

Monthly time-weighted all-hours mean GHI in W/m², including nights. Not daylight-only GHI, POA, or irradiation in kWh/m².

temperature_c

All tables

Monthly time-weighted mean ambient temperature in °C. Convert CMIP6 tas from kelvin before use.

days

All tables

Actual integer days in each native-calendar month. Gregorian, noleap, all_leap and 360_day year patterns are supported explicitly.

dc_yield_kwh_kwp

Measured only

Monthly total DC energy divided by qualified DC STC capacity, in kWh/kWp. Not mean daily yield, AC output, or energy in MWh.

For regular interval inputs, integrate power over actual elapsed interval hours to obtain energy, and use time-weighted means for weather. Check complete interval coverage before monthly aggregation: a monthly sum from a partial month is not a qualified monthly total. This API cannot detect missing hours hidden inside a caller’s finite aggregate. It does not fill gaps or scale partial months to full months. Valid zero-yield months remain zero. Zero annual measured yield and zero model-GHI monthly centroids are rejected because normalization is undefined.

The existing station connectors can supply measured weather; actual DC energy must come from qualified plant measurements. The existing CMIP6 connector retrieves native files and provenance. This PR does not add a NetCDF parser, spatial interpolation, automatic station selection or hourly disaggregation. Reduce native model calendars to year/month labels and keep their days explicitly. Never coerce a 360_day model into Gregorian timestamps.

Use the same climate model, member, variable definitions, grid/location and preprocessing for historical calibration and its future scenario. Keep each model/member separate, and preserve acquired dataset identifiers, versions, coordinates, hashes, calendar and source citations alongside model.to_dict(). The method cannot verify provider lineage from a numeric table.

Calculations and conventions#

For calendar month m, calculate a mean daily yield from each measured month’s total divided by its measured-calendar days. With bars denoting the average across calibration years, the invariant response is:

c[m] = mean(measured monthly DC yield / measured days)[m]
       / mean(raw historical model GHI)[m]
Y_irr[t] = c[month(t)] * raw scenario GHI[t] * scenario days[t]

This captures the plant’s seasonal response while retaining climate-model GHI variability. The raw model GHI is intentionally used here: its baseline bias is already absorbed in c. Applying GHI site adaptation again to this yield response would correct that bias twice. The separate GHI adaptation supplies the Ross temperature model instead.

G_site[t] = a_G * G_model[t] + b_G
T_site[t] = a_T * T_model[t] + b_T
T_module[t] = T_site[t] + ross * G_site[t]
T_reference[m] = mean(a_T * historical model T + b_T)[m]
                 + ross * mean(a_G * historical model GHI + b_G)[m]
f_temperature[t] = 1 + gamma * (T_module[t] - T_reference[month(t)])
f_power[t] = 1 - degradation_per_year * (age_at_start_years + operating_year[t] - 1)

The temperature multiplier is an incremental, first-order adaptation: the measured yield calibration already includes historical thermal losses. It is not an additional absolute STC loss, nor a claimed exact reproduction of the unpublished normalization behind Figure 7.17. Unchanged reference weather gives a multiplier of one. The historical reference uses the same fitted weather maps as the future projection: seasonal regression residuals must not masquerade as future temperature changes. Monthly averages and GHI reproduce the chapter’s coarse approach; this is not an irradiance-weighted hourly module-temperature model.

An existing plant’s historical aging and operating losses remain embedded in its measured yield. Supply an explicitly justified age_at_start_years relative to the calibrated yield reference, or a relative power-retention curve, to avoid applying losses already present in the baseline twice. There is no automatic weather-driven acceleration of degradation, replacement policy or rejuvenation.

Four outputs keep independent and combined effects distinguishable:

Output

Monthly specific DC yield

irradiance_kwh_kwp

Y_irr

irradiance_temperature_kwh_kwp

Y_irr * f_temperature

irradiance_power_kwh_kwp

Y_irr * f_power

combined_kwh_kwp

Y_irr * f_temperature * f_power

The printed equations 7.2-7.4 omit a clear -1 for signed gain/loss and put an integral together with a year count in the denominator. The surrounding text describes total energy compared with an annual P-value times years, and the figures show zero-centered signed percentages. Therefore this implementation exports both 100 * cumulative_yield / (annual_baseline * years) and 100 * (cumulative_yield / (annual_baseline * years) - 1). The denominator is not multiplied by years twice. A constant baseline yields 100% ratio and 0% gain.

P10/P50/P90 are empirical nonexceedance quantiles (Q0.10/Q0.50/Q0.90, linear interpolation) of complete measured annual DC yields. This choice agrees with the ordering of the thesis’s comparisons; it is explicitly not the finance convention in which P90 is a low energy yield exceeded 90% of the time. These quantiles describe historical interannual variation only. They do not combine model spread, scenario likelihood, regression uncertainty or outages, and are not bankable P90 guarantees. Different SSPs are never pooled or assigned probabilities.

The method allows future climate outside historical support, flags it by month and counts flagged months per year. The GHI check uses calendar-month bounds; temperature uses the overall adaptation range. These are diagnostic ranges, not proof of future accuracy. Negative adapted irradiance, nonpositive thermal multipliers, invalid power retention and incomplete inputs raise errors instead of being silently clipped or replaced. There are no random seeds because all calculations are deterministic.

Usage#

from energy_methods.ccya import fit_ccya

# Tables follow the schema above; select the calibration overlap explicitly.
model = fit_ccya(measured_monthly, historical_model_monthly)
result = model.project(
    future_model_monthly,
    scenario="model/member/ssp245/grid/version",
    degradation_per_year=0.005,  # 0.5%/year, not 0.5
)
monthly = result.monthly
annual = result.annual
provenance = result.metadata  # Includes the fitted model and REFERENCE_IDS.

For a nonlinear path, pass power_factors=pd.Series(...) with exactly the projection’s monthly index and values in [0, 1]; leave the linear rate and age at zero. For a scenario ensemble, call project separately for each scenario and each member. Use the matching fitted historical model for each member.

Run the deterministic synthetic example, which prints model provenance and 25/50-year comparisons for two artificial climate paths and the four degradation rates illustrated in the thesis:

uv run --no-sync python methods/examples/ccya_synthetic.py
uv run --no-sync pytest methods/tests/test_ccya.py

The example makes no real-site or real-SSP claim and downloads no data. Tests independently check affine bias correction, a known warming response, irradiance-induced Ross heating, constant-baseline identity, cumulative linear degradation, nonlinear power factors, percentile direction, complete rolling windows, calendar handling and invalid-input rejection.