Methodology

How this simulation works — every model equation with its parameters defined, the statistical processes used and where, and the exact data taken from each API (variables, native units, conversions). This section reflects the current implementation and is updated whenever the model changes. Section numbers match the internal physics reference (§1–§19).

Overview▾

The simulation runs an 8,760-hour (full year, hourly) model at a user-specified location, vectorised over the whole year at once. It combines real-world weather data with physics-based models to estimate yield accounting for irradiance, temperature, wind, soiling, snow, and long-term degradation. Two configurations are always evaluated in parallel: a two-axis tracking system (the product) and a fixed-tilt reference system (the industry-standard comparison).

Two data paths exist: the primary path uses measured PVGIS typical-year weather (§2); when that is unavailable (open ocean, beyond 73°N/56°S, API failure) a clear-sky fallback runs instead and results are flagged as optimistic. Each section below gives the governing equations, defines every symbol in a "where" table, and states its data source and any statistical processing. Hover any symbol inside an equation for its name and units; keep holding the hover for its full glossary entry; click it to jump to the glossary.

1. Solar Geometry▾

Concept — the seasons are a tilt, not a distance

The Earth's axis is tilted 23.45° from its orbital plane and — like a spinning gyroscope — keeps pointing at the same spot in space all year while the planet circles the Sun (Fig. 1.1). So for half the orbit the southern hemisphere leans toward the Sun and for the other half it leans away. That lean is the entire cause of the seasons: the Earth–Sun distance varies only ~3%, and the Earth is in fact closest to the Sun in early January — midsummer in the southern hemisphere.

Solar declination δ compresses the lean into one number: the latitude at which the sun is directly overhead at noon. It swings sinusoidally between +23.45° (June solstice) and −23.45° (December solstice), crossing zero at the equinoxes — equation (1.1). δ is the calendar's contribution to the sun's position; every equation after it takes δ as an input.

SunNNNNδ = +23.45°June solsticeN leans toward SunDecember solsticeN leans away · δ = −23.45°March equinox · δ = 0September equinox · δ = 0
Fig. 1.1 — The Earth's axis keeps a fixed direction in space (23.45° from the orbit's vertical) while the planet circles the Sun, so each hemisphere leans sunward for half the year. Declination δ is the angle between the Sun's rays and the equatorial plane: +23.45° at the June solstice, zero at the equinoxes, −23.45° in December. Original figure; geometry after Masters (2013).

Concept — sundials disagree with clocks — solar time

Clock time is a convenient fiction: it ticks at a perfectly uniform rate and an entire time zone shares one value. The real sun does neither, and the geometry needs the real sun, so two corrections convert clock time to true solar time (1.4).

First, longitude: within a zone, a site east or west of the zone's reference meridian sees solar noon earlier or later by 4 minutes per degree. Second, the equation of time (1.3): even at the reference meridian a sundial drifts against the clock over the year, because the Earth's orbit is slightly elliptical (it moves faster near its January closest approach — Kepler's second law) and because the axial tilt projects the sun's motion unevenly onto the equator. The two effects stack into the lopsided ±16-minute wobble in Fig. 1.2.

-15-10-5+5+10+150JFMAMJJASONDsundial − clock (minutes) · + sundial ahead, − sundial behindmid-Feb ≈ −14 minearly Nov ≈ +16 min
Fig. 1.2 — The equation of time — how far a sundial runs ahead of (+) or behind (−) a uniform clock through the year, from the Spencer (1971) series used by the model. The two causes: Earth's orbit is slightly elliptical (it moves faster near perihelion in January), and the axial tilt projects the Sun's motion unevenly onto the equator. The swing spans about half an hour: ≈ −14 min in mid-February to ≈ +16 min in early November.

Concept — the Earth is a 15°-per-hour clock — the hour angle

Once solar time is known, the hour angle H (1.5) restates it as pure geometry: the Earth turns 360° in 24 hours — 15° per hour — so "three hours before solar noon" and "the sun's meridian is 45° away" are the same statement (Fig. 1.3). H = 0 at solar noon; this model counts mornings negative and afternoons positive. H is the clock's contribution to the sun's position, as δ was the calendar's.

N poleSunsolar-noon meridianyour meridian — 9:00 solar timeH = −45°15:00 solar time — H = +45°Earth turns 15°/hourviewed from above the North Pole
Fig. 1.3 — The hour angle H, viewed from above the North Pole. The Earth turns 15° per hour, so H measures how far your meridian (your line of longitude) is from the solar-noon position — a clock face written in degrees. This model counts mornings negative and afternoons positive: at 9:00 solar time the Sun is still 45° 'ahead', so H = −45°. Original figure; concept after Masters (2013).

Concept — two numbers pin the sun to your sky

From δ (the season), H (the time of day) and latitude L (where you stand), two spherical-trigonometry identities give the sun's coordinates in your local sky (1.6, 1.7): the elevation β — its angle above the horizon — and the azimuth φs — its compass bearing (Fig. 1.4). Sign convention: the model measures φs from due north — 0° is north, ±180° due south — and takes its sign from the hour angle: negative bearings lie east of north (mornings), positive lie west (afternoons). This is how a ship's lookout reports a contact: a bearing and a height. Everything downstream consumes exactly these two angles — the two-axis tracker aims with them directly (§5), and the fixed-tilt panel compares them against its own fixed orientation to get its incidence angle (§6).

horizon planeNESWobserver / panelSunβelevationφₛazimuth (from north)
Fig. 1.4 — Two numbers pin the Sun to your sky, the way a lookout reports a ship: a compass bearing and a height. The azimuth φₛ is the bearing of the Sun's ground position measured from north; the elevation β is the angle from the horizon up to the Sun. The model computes both for every daylight hour; everything downstream — tracker aim, panel incidence — is built from them. Original figure.

Concept — air mass — the long slant path

The lower the sun, the more atmosphere the beam must cross before reaching the panel (Fig. 1.5): straight overhead is one atmosphere's worth (m = 1); at 30° elevation the slant doubles it; at 10° it is ~5.6×. More air in the path scatters and absorbs more of the beam — the same geometry that makes a low sun dim and red. The Kasten–Young (1989) formula (1.8) refines the naive 1/sin β estimate near the horizon, where the Earth's curvature caps the path at ≈38 atmospheres rather than the infinity a flat-slab model would give. Air mass is consumed only by the ASHRAE clear-sky fallback (§2.2) — measured PVGIS irradiance already contains the real attenuation.

top of atmosphere (schematic)paneloverhead: m = 130° above horizon: m ≈ 230°10°: m ≈ 5.6
Fig. 1.5 — Air mass m — the length of the light's slanting path through the atmosphere, relative to straight overhead. At 30° elevation the path is about twice as long (m ≈ 2); at 10° it is ~5.6×. The same geometry explains why the low sun is dim and red: more air scatters more light out of the beam. Used only by the ASHRAE clear-sky fallback (§2) — measured PVGIS data has real attenuation built in. Original figure.
(1.1) Declination (thesis 1)
δ=23.45∘ sin⁡ ⁣(360(284+n)/365) [∘]{\htmlData{sym=delta}{\delta}} = 23.45^\circ\,\sin\!\big(360(284 + {\htmlData{sym=nday}{n}})/365\big)\ [^\circ]
(1.2) Day angle (Spencer 1971)
B=360(n−1)/365 [∘]{\htmlData{sym=Bday}{B}} = 360({\htmlData{sym=nday}{n}} - 1)/365\ [^\circ]
(1.3) Equation of time (Spencer 1971)
EoT=229.18 ( 0.000075+0.001868cos⁡B−0.032077sin⁡B−0.014615cos⁡2B−0.04089sin⁡2B ) [min]\begin{aligned} {\htmlData{sym=EoT}{\text{EoT}}} = 229.18\,(\, &0.000075 + 0.001868\cos {\htmlData{sym=Bday}{B}} - 0.032077\sin {\htmlData{sym=Bday}{B}} \\ &- 0.014615\cos 2{\htmlData{sym=Bday}{B}} - 0.04089\sin 2{\htmlData{sym=Bday}{B}}\,)\ \text{[min]} \end{aligned}
(1.4) Solar time (thesis 3.1)
tsolar=tlocal+(λ−λstd)/15+EoT/60 [h]{\htmlData{sym=t_solar}{t_{\text{solar}}}} = {\htmlData{sym=t_local}{t_{\text{local}}}} + ({\htmlData{sym=lambda}{\lambda}} - {\htmlData{sym=lambda_std}{\lambda_{\text{std}}}})/15 + {\htmlData{sym=EoT}{\text{EoT}}}/60\ [\text{h}]
(1.5) Hour angle (thesis 2, sign flipped)
H=15 (tsolar−12) [∘]{\htmlData{sym=Hang}{H}} = 15\,({\htmlData{sym=t_solar}{t_{\text{solar}}}} - 12)\ [^\circ]
(1.6) Solar elevation (thesis 5)
β=arcsin⁡ ⁣(sin⁡L sin⁡δ+cos⁡L cos⁡δ cos⁡H) [∘]{\htmlData{sym=beta}{\beta}} = \arcsin\!\big(\sin {\htmlData{sym=Lat}{L}}\,\sin{\htmlData{sym=delta}{\delta}} + \cos {\htmlData{sym=Lat}{L}}\,\cos{\htmlData{sym=delta}{\delta}}\,\cos {\htmlData{sym=Hang}{H}}\big)\ [^\circ]
(1.7) Solar azimuth (arccos form; thesis 6 equivalent)
φs=sgn⁡(H) arccos⁡ ⁣(sin⁡δ−sin⁡L sin⁡βcos⁡L cos⁡β) [∘]{\htmlData{sym=varphi_s}{\varphi_s}} = {\htmlData{sym=sgnf}{\operatorname{sgn}}}({\htmlData{sym=Hang}{H}})\,\arccos\!\Bigg(\dfrac{\sin{\htmlData{sym=delta}{\delta}} - \sin {\htmlData{sym=Lat}{L}}\,\sin{\htmlData{sym=beta}{\beta}}}{\cos {\htmlData{sym=Lat}{L}}\,\cos{\htmlData{sym=beta}{\beta}}}\Bigg)\ [^\circ]
(1.8) Air mass (Kasten–Young)
m=1sin⁡β+0.50572 (β+6.07995)−1.6364{\htmlData{sym=mass}{m}} = \dfrac{1}{\sin{\htmlData{sym=beta}{\beta}} + 0.50572\,({\htmlData{sym=beta}{\beta}} + 6.07995)^{-1.6364}}
SymbolMeaningUnitsValue / source
nDay of year1–365derived from the hour index (hour 0 = 1 Jan, 00:30)
BDay angle (Spencer)°360(n−1)/365 — not the ASHRAE coefficient B in §2
EoTEquation of time — orbital eccentricity/obliquity clock driftminSpencer (1971) series; drift up to ±16 min
tlocalLocal clock time (mid-hour convention: +0.5 h)hhour index; TMY path adds round(λ/15) h UTC offset first
λSite longitude (east positive)°user input
λstdStandard meridian of the site's time zone°round(λ/15) × 15
tsolarTrue solar time (sun crosses the meridian at 12:00)hthesis eq 3.1
HHour angle — angular distance from solar noon°15° per hour; negative mornings, positive afternoons
LSite latitude (north positive)°user input
δSolar declination — seasonal tilt of the sun's path°±23.45° over the year (thesis eq 1)
βSolar elevation above the horizon°thesis eq 5; β ≤ 0 = sun below horizon
φsSolar azimuth from north°arccos form; negative east, positive west
sgnSignum function — the sign of its argument (+1, −1, or 0)—applies the hour angle's sign to the azimuth in (1.7)
mRelative air mass — atmosphere path-length ratio—Kasten–Young; used only by the ASHRAE fallback (§2)

Worked example — Perth (L = −32°, λ = 115.9°E)

21 June (n = 172): δ = +23.45° — the northern hemisphere leans sunward. Noon elevation β = 90° − |L − δ| = 90° − 55.4° = 34.6°: the midwinter sun climbs barely a third of the way up the sky, due north. 21 December (n = 355): δ = −23.45°, noon β = 81.4° — nearly overhead.

Clock → solar: Perth lies 4.1° west of its zone meridian (120°E), so solar noon runs ≈17 minutes behind clock noon; adding the equation of time for 21 June (−1.3 min) puts solar noon at ≈ 12:18 clock time. At 9:00 solar time the hour angle is H = 15 × (9 − 12) = −45°. (All values computed with the model's own functions.)

Notation — why L for latitude: the solar-geometry literature splits into two self-consistent families. Masters (2013) — and the thesis this model descends from — write latitude as L and reserve φ for the azimuths (φs solar, φp panel). Duffie & Beckman write latitude as φ but then move the azimuths to γ — a symbol already taken here by the module temperature coefficient (§9). Mixing the two (φ for latitude alongside φs for azimuth) would overload φ within single equations, so this model follows Masters/the thesis throughout: latitude is always L, and any φ is an azimuth. Longitude keeps its geodetic symbol λ, which collides with nothing.

Provenance notes: the declination phase (284+n) is identical to the thesis's (n−81) form. The equation of time is the Spencer (1971) five-term Fourier series — more precise than the thesis's three-term Masters (2013) form (they agree to within about half a minute). The azimuth uses the standard arccos formula (1.7); the thesis computes the same angle through its sine form with a quadrant check (thesis eqs 6, 6.1): sin φs = cos δ · sin H / cos β, with the quadrant resolved by the inequality cos H ≥ tan δ / tan L ⟹ |φs| ≤ 90°, else |φs| > 90° — needed because the arcsine is ambiguous between φs and 180° − φs. The arccos form needs no check (the sign comes from the hour angle), which is why the code prefers it; the thesis's east/west sign convention is also opposite (positive east).

Numerical guards: arcsine/arccosine inputs are clamped to [−1, 1]; the azimuth denominator carries +10⁻¹⁰ to avoid dividing by zero with the sun at the zenith; the air-mass elevation is clipped to ≥ 0.1° to avoid an infinite path length at the horizon.

UTC alignment (TMY path): PVGIS timestamps are UTC, so the model shifts the hour of day by round(λ/15) hours before the geometry runs. Without this, midday irradiance would be paired with midnight sun positions — a ~3× yield error.

2. Irradiance — Weather Data▾

Concept — three kinds of light

Sunlight arrives at a panel by three routes (Fig. 2.1). Direct beam (DNI) — parallel rays straight from the solar disc; the only component with a direction, so the only one a tracker can chase and the only one that casts a shadow. Diffuse (DHI) — light scattered by air molecules and cloud, arriving from the whole sky dome; on an overcast day it is all there is. Ground-reflected — the fraction ρ (the albedo) of total sunlight that bounces off the ground onto the panel face. The three are modelled separately because geometry treats them differently: beam responds to panel orientation hour-by-hour (§5, §6), while diffuse and ground light depend only on how much sky and ground the panel can "see".

paneldirect beam (DNI)diffuse (DHI)ground-reflected (ρ × GHI)
Fig. 2.1 — Sunlight reaches a panel three ways. Direct beam (DNI): parallel rays straight from the solar disc — the component trackers chase, and the only one that casts a shadow. Diffuse (DHI): light scattered by air and cloud, arriving from all over the sky dome. Ground-reflected: the fraction ρ (albedo) of total horizontal irradiance (GHI) that bounces off the ground onto the panel face. GHI is the sum of beam and diffuse on a horizontal surface. Original figure.

Concept — a measured year vs a cloudless ceiling

The primary data path uses measured satellite irradiance: real clouds, real haze, real weather, hour by hour, assembled into a statistically typical year (the TMY — full explainer in §16). The ASHRAE fallback is a different kind of object entirely: a physics formula for a cloudless sky — a ceiling, not a forecast, like quoting a road's speed limit instead of its actual traffic. That is why the two are never averaged: one describes what the sky did, the other what it could do at best.

Primary: PVGIS TMY (Typical Meteorological Year) from the EU Joint Research Centre (JRC). A statistically representative year assembled from the 2005–2020 (16-year) SARAH-3 / ERA5 satellite and reanalysis record (~5 km grid) — each calendar month is selected independently as the most distribution-representative of its 16 candidates (see the TMY explainer in §16). Provides hourly direct normal irradiance (DNI, field Gb(n)), diffuse horizontal irradiance (DHI, Gd(h)), global horizontal irradiance (GHI, G(h)) — all W/m² — and 2 m air temperature (T2m, °C). Cloud cover is embedded in the measured values. Coverage: 73°N to 56°S land areas.

Fallback: ASHRAE clear-sky model — used only when PVGIS is unavailable (extreme latitudes, API failure, open ocean). A theoretical upper bound with no cloud attenuation; results are flagged to the user as optimistic estimates. Three coefficients drive it, defined individually in the table below: A, the apparent extraterrestrial flux (top-of-atmosphere solar intensity, varying with Earth–Sun distance); Batm, the atmospheric extinction coefficient (total optical depth per unit air mass, all constituents — a broadband climatological quantity, not the aerosol-only AOD550 used for soiling in §10); and C, the sky diffuse factor (diffuse irradiance as a fraction of DNI). All three are looked up from the ASHRAE Handbook's monthly table — 12 values per coefficient, hardcoded as _ASHRAE_A/B/C in solar_model.py. The two irradiance sources are never averaged — they describe different things (measured weather vs a cloudless ceiling); blending them would be physically meaningless.

(2.1) Direct normal — ASHRAE clear-sky (thesis 12)
DNI=A exp⁡(−Batm m) [W/m2]{\htmlData{sym=DNI}{\text{DNI}}} = {\htmlData{sym=A_ash}{A}}\,\exp(-{\htmlData{sym=B_ash}{B_{\text{atm}}}}\,{\htmlData{sym=mass}{m}})\ [\text{W/m}^2]
(2.2) Diffuse (thesis 16)
Id=C ⋅ DNI [W/m2]{\htmlData{sym=I_d}{I_d}} = {\htmlData{sym=C_ash}{C}}\,\cdot\,{\htmlData{sym=DNI}{\text{DNI}}}\ [\text{W/m}^2]
(2.3) Global horizontal (thesis 13 + 16)
GHI=DNI sin⁡β+Id [W/m2]{\htmlData{sym=GHI}{\text{GHI}}} = {\htmlData{sym=DNI}{\text{DNI}}}\,\sin{\htmlData{sym=beta}{\beta}} + {\htmlData{sym=I_d}{I_d}}\ [\text{W/m}^2]
(2.4) Source gate — measured data preferred, clear-sky only as fallback
{DNI, Id, GHI}={PVGIS TMY measured valuesnormalASHRAE clear-sky (eqs 2.1–2.3)fallback\{{\htmlData{sym=DNI}{\text{DNI}}},\ {\htmlData{sym=I_d}{I_d}},\ {\htmlData{sym=GHI}{\text{GHI}}}\} = \begin{cases} \text{PVGIS TMY measured values} & \text{normal} \\[2pt] \text{ASHRAE clear-sky (eqs 2.1–2.3)} & \text{fallback} \end{cases}
(2.5) Ground-reflected (thesis 18 form) — either path
Ir=ρ ⋅ GHI [W/m2]{\htmlData{sym=I_r}{I_r}} = {\htmlData{sym=rho}{\rho}}\,\cdot\,{\htmlData{sym=GHI}{\text{GHI}}}\ [\text{W/m}^2]
SymbolMeaningUnitsValue / source
DNIDirect (beam) irradiance at normal incidenceW/m²PVGIS Gb(n), or A·exp(−B·m) on the fallback
IdDiffuse horizontal irradiance (sky-scattered light)W/m²PVGIS Gd(h), or C·DNI on the fallback
GHIGlobal horizontal irradiance = beam projection + diffuseW/m²PVGIS G(h), or DNI·sin β + Id
IrGround-reflected irradianceW/m²ρ × GHI
AApparent extraterrestrial flux — top-of-atmosphere intensity (fallback only)W/m²looked up: ASHRAE Handbook monthly table (_ASHRAE_A, 12 values)
BatmAtmospheric extinction — total optical depth per unit air mass (fallback only)—looked up: ASHRAE Handbook monthly table (_ASHRAE_B); broadband total, not the aerosol-only AOD550 of §10
CSky diffuse factor — diffuse as a fraction of DNI (fallback only)—looked up: ASHRAE Handbook monthly table (_ASHRAE_C)
mAir mass (fallback only)—Kasten–Young, from §1
ρGround albedo — fraction of GHI reflected back onto the panel—site ground-surface input, table below; default 0.20

Ground albedo ρ is the reflectance of the surface around your site — you choose it directly (accepted range 0.05–0.9; if you don’t, the model uses 0.20, typical of grass or mixed ground). Preset values come from the IEA PVPS Task 13 bifacial report (2021), Table 7, after Iqbal (1983), unless noted:

SurfaceGround albedo (ρ)Published range / source
Grass / mixed vegetation0.200.15–0.26
Black soil0.100.08–0.13
Clay soil0.200.16–0.23
Sand0.350.21–0.60
Asphalt0.140.09 (new) – 0.18 (weathered)
White membrane roof0.55measured, lightly soiled — Marion et al. (2017)
White concrete0.50Sun et al. (2018)

Snow-covered ground uses ρ = 0.65: in months whose snowfall passes the §11 gate, the value above is lifted to max(ρ, ρsnow) inside the hourly simulation itself — every ground-reflected term, front and rear, both configurations, and the §7 tilt sweep. The §11 "albedo bonus" figure is an attribution estimate of that embedded effect, not an addition.

Seasonal albedo (automatic for natural surfaces): real ground gets darker and lighter through the year — vegetation greens up and dries off, soil wets and dries. When the chosen surface is regional land cover (grass, soils, sand), each snow-free month automatically uses your site albedo scaled by the satellite’s seasonal pattern for the surrounding cell: ρ × (satellite month ÷ satellite snow-free mean), days-weighted so the average over those months still equals your chosen level exactly (results stay inside the 0.05–0.9 accepted range). The satellite contributes only the shape: its ~50 km cell (NASA POWER ALLSKY_SRF_ALB, 2001–2020 climatology) measures regional land cover, while the ground your panel actually sees is yours — so your surface choice keeps setting the level, and the cell’s annual value is shown beside the picker as a sanity check. There is deliberately no on/off switch, and two cases deliberately stay static: built surfaces (asphalt, membrane roof, concrete), because a roof does not follow the region’s vegetation cycle — applying the shape there would be wrong physics, not extra precision — and custom albedo, because an unidentified surface supports no seasonal claim. Months passing the §11 snow gate are excluded from the pattern for every surface: the satellite’s winter brightening is snow, which the max(ρ, ρsnow) lift above already applies — scaling the shape there would count the same snow twice.

Spectral factor (PV-effective albedo): a cell only converts light inside its absorption window (silicon: roughly 300–1200 nm), while broadband albedo weights the ground’s reflectance by the whole solar spectrum. Where the two disagree the model corrects for it: every ground-reflected term — Ir on both configurations’ fronts, both bifacial rears (§17), and the §7 tilt sweep, never sky or beam light — is multiplied by a factor k specific to the surface–technology pairing. The classic case is vegetation’s red edge: green grass reflects weakly below ~700 nm but strongly in the near-infrared that silicon still converts, so broadband ρ = 0.20 undercounts what a silicon module actually receives (k = 1.17); a CdTe module, whose window closes earlier, sees the opposite (k = 0.79). Factors are exact ratios of spectrally-weighted to broadband albedo from Tonita et al. (2024), Table 1:

Surfacec-Si (mono/poly)HJTCIGSCdTe
Grass / mixed vegetation1.1671.1251.1670.792
Black / clay soil0.9470.9470.9470.789
Sand1.0001.0001.0000.879
White concrete0.9661.0001.0000.931
Asphalt · white membrane1.000 (assumed)1.000 (assumed)1.000 (assumed)1.000 (assumed)
Snow (§11 gated months)1.1061.1061.1061.141

Snow-covered months use the snow row on the brighter of the two PV-effective grounds, max(ρ·ksurface, ρsnow·ksnow) — snow is identified ground, so its factor applies even when the surface is unknown. Asphalt and white membrane carry no correction as a named assumption (no measured spectrum for them in the sources opened); a custom albedo gets none either — the model makes no spectral claims about ground it cannot identify. The §11 albedo-bonus figure remains a broadband attribution estimate; the simulated yields carry the factor. PV-effective albedo is capped at 0.98 (reflectance cannot exceed 1). For scale: fixed-tilt systems over vegetation showed rear spectral impacts up to 1.20 in year-long field measurements (Riedel-Lyngskær et al. 2022) — omitting spectral albedo moves rear-side irradiance estimates by up to ±13% (Tonita et al. 2024).

3. Ambient Temperature & Lapse-Rate Correction▾

Hourly ambient temperature comes from the PVGIS TMY (T2m). PVGIS reports at its ~5 km grid-cell elevation, which can differ substantially from the actual site (a mine at 1,200 m in a grid cell averaging 400 m). Temperature is corrected with the ICAO standard atmospheric lapse rate:

Concept — the mountain is not its grid cell

Air cools as it rises and expands — the same reason mountaintops hold snow in summer. Through the lower atmosphere the decline is remarkably linear, and the ICAO standard atmosphere fixes it at Γ = 6.5 °C per 1,000 m (about 1 °C per 154 m). A weather dataset can be perfectly accurate at its own elevation and still wrong for your site: PVGIS reports one temperature for a ~5 km cell at the cell's mean elevation, so a ridge site 800 m above that mean is really ≈5.2 °C cooler than the data says (Fig. 3.1). Left uncorrected, the error flows through both heat penalties — overstating the instantaneous derating (§9) and the 25-year degradation rate (§12) for high sites, and understating them for valleys.

your site — 1,200 mPVGIS grid-cell meanelevation — 400 mΔz = 800 m→ −5.2 °C (Γ = 6.5 °C/km)
Fig. 3.1 — A PVGIS grid cell is ~5 km wide and reports temperature at its mean elevation. A mine site on a ridge at 1,200 m inside a cell averaging 400 m is 800 m higher than the data thinks — and at the ICAO lapse rate Γ (6.5 °C per 1,000 m) that is 5.2 °C cooler in reality. The correction matters twice: cooler cells produce more now (§9) and degrade slower over 25 years (§12). Original figure.
(3.1) ICAO lapse-rate correction
Tamb=TPVGIS−0.0065 ⋅ (elevsite−elevgrid) [∘C]{\htmlData{sym=T_amb}{T_{\text{amb}}}} = {\htmlData{sym=T_PVGIS}{T_{\text{PVGIS}}}} - 0.0065\,\cdot\,({\htmlData{sym=elev_site}{\text{elev}_{\text{site}}}} - {\htmlData{sym=elev_grid}{\text{elev}_{\text{grid}}}})\ [^\circ\text{C}]
SymbolMeaningUnitsValue / source
TambCorrected ambient temperature (feeds §8)°Coutput
TPVGISPVGIS TMY 2 m air temperature°Chourly T2m; Open-Meteo ERA5 daily means (2018–2022) on the fallback path
ΓICAO standard lapse rate°C/m0.0065 — about 1 °C per 154 m
elevsiteActual site elevationmOpen-Topo-Data ASTER (~30 m); Open-Elevation SRTM fallback
elevgridReference elevationmPVGIS grid-cell metadata; 0 m (sea level) on the clear-sky path

The correction is only applied when the elevation difference is at least 1 m (the value is still reported for display). The error it removes propagates through both the Faiman model (instantaneous derating) and the Arrhenius model (25-year degradation) — high-altitude sites run cooler and degrade slower.

Fallback path only: the 12 monthly means are expanded to 8,760 hours by circular interpolation between mid-month anchor points plus a ±3 °C diurnal cosine peaking at 14:00. The TMY path uses the measured hourly series directly. If ERA5 is also unavailable, a latitude-based estimate is used and flagged in the data-source string.

4. Shading▾

Shading arrives in two regimes with different physics. Far-field obstructions — terrain, distant buildings, tree lines — screen the whole collector uniformly: a pure irradiance effect. Near-field obstructions — fences, walls, roof furniture — shade parts of the module, and partial shade is an electrical effect: a crystalline-silicon substring answers its worst cell, so a small shadow can cost a disproportionate share of power. Both regimes are optional blocks on the API request (horizon, near_objects), resolved once in the router and threaded to both simulation paths; with neither block the run is bit-identical to an unshaded one. (An input UI is deliberately deferred — today these inputs arrive through the API only.)

Concept — far vs near — two kinds of shadow

A distant hill and a close fence can throw the same-sized shadow and do completely different things to yield. The hill is far enough away that it screens the module — and every module like it — uniformly: some beam hours simply never arrive, and a slice of the sky is hidden. That is the far-field regime — geometry only, no electrics.

The fence is close enough that its shadow lands on part of the module — some cell strips dark, their neighbours in full sun. That mismatch is what makes partial shade expensive: the module's bypass diodes answer the worst-shaded strip of each substring, so a shadow covering a fraction of a substring can take the substring's whole beam contribution. That is the near-field regime — geometry plus the §9.3 electrical layer.

The horizon line h(φ): the far-field input is a skyline — (compass azimuth, elevation) points interpolated piecewise-linearly with wrap-around. A profile peaking under 2° is insignificant (the PVsyst threshold) and resolves to exactly no horizon. For beam, each hour keeps the fraction of the hour the sun spends above the skyline — four 15-minute substeps, so the hour the sun crosses the skyline goes fractional instead of all-or-nothing (eq 4.1, Fig 18.1). The gate is applied once to the shared direct normal irradiance, so every beam term — both configurations' fronts, the §7 tilt sweep, the §10.3 αr tables and the §17 bifacial rear beam — sees the same sun. A fully clear hour keeps exactly 1, a fully blocked one exactly 0.

Concept — why sub-hour, not ON/OFF

The simpler gate — compare the mid-hour sun position against the skyline and take or leave the whole hour — carries a hidden bias. Which way the crossing hour falls is set by the sampling phase: the fixed offset between the local-time grid and solar noon, drifting daily with the equation of time. That is a systematic, seasonally clustered error, not zero-mean noise. On a 40°-wall stress case, where clear-sky noon symmetry requires zero, ON/OFF gating produced a spurious 1.39-percentage-point annual-DNI asymmetry between mirrored east and west walls. Four substeps cut it to 0.19 pp — within 0.02 pp of a near-continuous 60-substep fraction, so the gate itself is converged; what remains is the hourly resolution of the irradiance data itself, irreducible at the model's (and the TMY's) native timestep.

10°20°30°40°50°sun elevation [°]E · 90°S · 180°W · 270°compass azimuth φsolar noonclear of the skyline — fvis = 1the crossing hour in 15-min substeps —2 of 4 above the skyline → fvis = ½falls behind the skyline ~15:40the hill takes the last hour —unblocked, the sun sets ~16:50h(φ) = 30°
Fig. 4.1 — The skyline h(φ) against the sun's daily path. Hours fully above the skyline keep fvis = 1 exactly, hours fully behind it 0; the crossing hour keeps the fraction of its four 15-minute substeps spent above the skyline — here 2 of 4, so fvis = ½ — instead of an all-or-nothing flip at the mid-hour sun position (the crossing-fraction refinement of the PVsyst far-shadings gate, eq 4.1). Original figure.

Diffuse loses a constant slice of sky: the horizon's diffuse factor fhor is the cosine-weighted fraction of the plane-visible sky dome that survives the skyline (eq 4.2). Because the skyline does not move with the sun, the factor is constant per plane orientation — computed once per geometry (PVsyst's treatment, after Quaschning & Hanitsch 1998) — and multiplies the isotropic diffuse alongside the §6.3 Martín–Ruiz Fd. The tilt sweep carries one factor per candidate tilt, so the optimum tilt genuinely shifts under an asymmetric horizon (test-pinned: an equator-side wall moves it shallower); the tracker interpolates its hourly orientations from a precomputed 7° × 15° tilt × azimuth grid. A horizontal plane under a flat horizon at elevation e gives exactly cos2e — the analytic anchor the integration is tested against.

Near objects shade strips, not the module: each near object is a vertical rectangle — centre at compass azimuth and distance, width perpendicular to the line of sight, heights relative to the module centre. Every hour its corners are projected along the sun vector onto the (moving, for the tracker) module plane and clipped exactly against the module's strip grid — N substrings across the width × C cell strips along it, from the module's electrical block (a generic 3 × 20 grid without one). The result is a per-strip shade fraction fj (eq 4.3, Fig 18.2), collapsing per substring to a worst-strip fraction fworst,k and a mean fmean,k, and module-wide to fmean. The diffuse side is a silhouette factor fnear on the same cosine-weighted dome as the horizon — exact ray/rectangle, no polygon approximation — folded in multiplicatively with fhor (eq 4.4): conservative where horizon and objects mask overlapping sky. The same sampling-phase trap as the horizon gate applies one level down, so the strip fractions are computed at four quarter-hour sun positions per hour — averaged for the linear tiers (their multipliers are linear in the fractions, so fraction-averaging is exact), solved per substep and averaged as multipliers for the exact diode tier, whose firing response is nonlinear. Measured on mirrored east/west walls: a spurious 0.81-percentage-point annual-loss asymmetry at mid-hour sampling, 0.07 pp with the substeps — converged (8 substeps drift it ≤ 0.03 pp).

near objectvertical rectanglemoduleshadow edgecross-sectionproject + clipsubstrings k123unclipped projectionfjfworst,3 = 1 · diode conductsfront view — module face
Fig. 4.2 — Near objects shade parts of the module. The obstruction's corners project along the sun vector onto the module plane (left) and clip exactly against the strip grid — N substrings across, C cell strips along (3 × 6 drawn for 3 × 20) — giving a per-strip shade fraction fj (eq 4.3). Column 3 has fully shaded strips, so fworst,3 = 1: a crystalline-silicon substring answers its worst strip — the bypass diode conducts and the substring's beam is gone, however small the shaded area — while thin-film loses only the area share fmean. Original figure.

Concept — the worst strip sets the substring

Why two fractions? Because technologies respond to shade differently. A crystalline-silicon substring is only as strong as its worst cell strip: once a strip is fully shaded, its reverse bias opens the bypass diode and the substring's beam is gone — however small the shaded area. That is why the c-Si tiers key on fworst. A thin-film module is monolithic — no substrings to knock out — and loses beam in proportion to shaded area: fmean is its loss path exactly (the documented divergence; its reverse-bias reliability caveat is a system-design caveat, not a yield term).

Three electrical tiers, one multiplier: c-Si with the full datasheet electrical block runs an exact bypass-diode solve — each substring its own translated single-diode curve at Sk = diffuse + ground + (1 − fworst,k)·beam, the module characteristic their series sum with each substring floored at its diode voltage, the operating point the global maximum (ideal MPPT — real-MPPT multi-peak behaviour and hot-spot analysis are deferred). c-Si without the curve runs the PVsyst partition — every shaded beam watt-hour on a substring fully lost (PVsyst's "fraction for electrical effect" at 100%; an area-proportional fallback would contradict the doctrine's disproportion). Thin-film takes the exact area-proportional form. The c-Si solve is also topology-aware: declaring a half-cut module runs the twin topology — three series stages, each a pair of parallel half-cell substrings sharing one bypass diode (PVsyst model documentation), so shade that respects the half split keeps the clear twins producing (the measured row-shading tolerance of half-cell modules; Applied Sciences 2020, 10:2575). A back-contact module never fires its bypass diodes: its cells are built to break down softly near −5 V, so each hour it takes the cheaper of two strategies — back the string current off to the worst cell, or hold current and pay the ≈ 5 V drop per driven cell. One fully shaded cell therefore costs ≈ 13% (the breakdown drop against the module voltage), not its ~1.5% area share — but a whole shaded substring costs far less than a knockout (TÜV NORD / DAS Solar measurements, 2026). Shingled modules fall to the partition tier — their strip structure is not mapped (a named limitation). All tiers report through msh (eq 4.5): the ratio of shaded to unshaded curve power, so the curve's absolute level cancels and the linear datasheet path keeps driving the energy. The multiplier is computed on shaded-beam against unshaded-diffuse components — it carries only the knock-out nonlinearity, never double-counting the diffuse loss — and scales hourly power on both configurations, the thermal variants and the bifacial baselines, so every gain ratio stays shade-consistent.

What the Partial-Shading Response section shows, and the MPPT caveat: the dashboard's scenario numbers (one cell, one substring, bottom row) are instantaneous losses for your module class at a bright reference hour, computed from the module parameters alone. Everything above assumes the operating point finds the global maximum of the shaded power curve — ideal maximum power point tracking (MPPT). That is exactly what a micro-inverter or DC optimizer realises per module; a basic tracker can instead settle on a local peak of the multi-peak curve partial shade creates (on our validation rig, 627 hours a year carried multi-peak curves, and the worst peak held a median of only 35% of the available power). Modern string inverters run periodic global sweeps that recover much of this, but verified recovery fractions for shipping hardware are not available in the literature we reviewed — so tracker behaviour is deliberately not a yield term here, and the exposure is reported instead (multi-peak hours ride the near-field echo alongside the bypassed-hours count). Hot-spot risk is likewise reported as a technology class, never a wattage: breakdown voltage spans −4 V (back-contact, deliberately soft) to −49 V (TOPCon-class) across 2026 cell technologies, appears on no datasheet, and varies −39 to −48 V within a single production bin (EPJ Photovoltaics 2024/2026; Solar Energy Materials & Solar Cells 2024) — a per-cell dissipation figure from datasheet inputs would be false precision.

(4.1) Far-field beam — sub-hour visibility fraction (PVsyst far shadings)
DNIeff(t)=DNI(t) ⋅ fvis(t)fvis(t)=1K∑k=1K1[β(tk)>h(φs(tk))],K=4\begin{aligned} {\htmlData{sym=DNI}{\text{DNI}}}_{\text{eff}}({\htmlData{sym=t_time}{t}}) &= {\htmlData{sym=DNI}{\text{DNI}}}({\htmlData{sym=t_time}{t}})\,\cdot\,{\htmlData{sym=f_vis}{f_{\text{vis}}}}({\htmlData{sym=t_time}{t}}) \\ {\htmlData{sym=f_vis}{f_{\text{vis}}}}({\htmlData{sym=t_time}{t}}) &= \frac{1}{{\htmlData{sym=K_sub}{K}}}\sum_{{\htmlData{sym=k_substr}{k}}=1}^{{\htmlData{sym=K_sub}{K}}} \mathbb{1}\big[{\htmlData{sym=beta}{\beta}}({\htmlData{sym=t_time}{t}}_{\htmlData{sym=k_substr}{k}}) > {\htmlData{sym=h_hor}{h}}({\htmlData{sym=varphi_s}{\varphi_s}}({\htmlData{sym=t_time}{t}}_{\htmlData{sym=k_substr}{k}}))\big], \qquad {\htmlData{sym=K_sub}{K}} = 4 \end{aligned}
(4.2) Far-field diffuse — constant sky-view factor per orientation
fhor(Σ,φp)=∫α>h(φ)cos⁡θi dΩ∫α>0cos⁡θi dΩ,dΩ=cos⁡α dα dφ{\htmlData{sym=f_hor}{f_{\text{hor}}}}({\htmlData{sym=Sigma}{\Sigma}}, {\htmlData{sym=varphi_p}{\varphi_p}}) = \frac{\displaystyle\int_{{\htmlData{sym=alpha_dome}{\alpha}} > {\htmlData{sym=h_hor}{h}}({\htmlData{sym=varphi_dome}{\varphi}})} \cos{\htmlData{sym=theta_i}{\theta_i}}\,{\htmlData{sym=dOmega}{\mathrm{d}\Omega}}}{\displaystyle\int_{{\htmlData{sym=alpha_dome}{\alpha}} > 0} \cos{\htmlData{sym=theta_i}{\theta_i}}\,{\htmlData{sym=dOmega}{\mathrm{d}\Omega}}}, \qquad {\htmlData{sym=dOmega}{\mathrm{d}\Omega}} = \cos{\htmlData{sym=alpha_dome}{\alpha}}\,{\htmlData{sym=d_diff}{\mathrm{d}}}{\htmlData{sym=alpha_dome}{\alpha}}\,{\htmlData{sym=d_diff}{\mathrm{d}}}{\htmlData{sym=varphi_dome}{\varphi}}
(4.3) Near-field beam — per-strip shade fractions (exact projection + clip)
fj=shaded area of strip jstrip area,fworst,k=max⁡strips j∈kfjfmean,k=1C∑j∈kfj,fmean=1N∑k=1Nfmean,k\begin{aligned} {\htmlData{sym=f_j}{f_j}} &= \frac{\text{shaded area of strip } {\htmlData{sym=j_strip}{j}}}{\text{strip area}}, & {\htmlData{sym=f_worst_k}{f_{\text{worst},k}}} &= \max_{\text{strips } {\htmlData{sym=j_strip}{j}} \in {\htmlData{sym=k_substr}{k}}} {\htmlData{sym=f_j}{f_j}} \\ {\htmlData{sym=f_mean_k}{f_{\text{mean},k}}} &= \frac{1}{{\htmlData{sym=C_strips}{C}}}\sum_{{\htmlData{sym=j_strip}{j}} \in {\htmlData{sym=k_substr}{k}}} {\htmlData{sym=f_j}{f_j}}, & {\htmlData{sym=f_mean}{f_{\text{mean}}}} &= \frac{1}{{\htmlData{sym=N_sub}{N}}}\sum_{{\htmlData{sym=k_substr}{k}}=1}^{{\htmlData{sym=N_sub}{N}}} {\htmlData{sym=f_mean_k}{f_{\text{mean},k}}} \end{aligned}
(4.4) Near-field diffuse — silhouette factor, multiplicative with the horizon
fnear(Σ,φp)=1−∫silhouettecos⁡θi dΩ∫α>0cos⁡θi dΩ,Idiffuse∝fhor⋅fnear{\htmlData{sym=f_near}{f_{\text{near}}}}({\htmlData{sym=Sigma}{\Sigma}}, {\htmlData{sym=varphi_p}{\varphi_p}}) = 1 - \frac{\displaystyle\int_{\text{silhouette}} \cos{\htmlData{sym=theta_i}{\theta_i}}\,{\htmlData{sym=dOmega}{\mathrm{d}\Omega}}}{\displaystyle\int_{{\htmlData{sym=alpha_dome}{\alpha}} > 0} \cos{\htmlData{sym=theta_i}{\theta_i}}\,{\htmlData{sym=dOmega}{\mathrm{d}\Omega}}}, \qquad {\htmlData{sym=I_diffuse}{I_{\text{diffuse}}}} \propto {\htmlData{sym=f_hor}{f_{\text{hor}}}} \cdot {\htmlData{sym=f_near}{f_{\text{near}}}}
(4.5) Electrical tiers — shade fractions become a power multiplier
msh(t)={Pcurve(S1,…,SN)Pcurve(S,…,S)exact diode solve (c-Si, curve)1−SbeamS fworst‾PVsyst partition (c-Si, no curve)1−SbeamS fmeanarea-proportional (thin-film)Sk=Sdiff+Sgnd+(1−fworst,k) Sbeam\begin{aligned} {\htmlData{sym=m_sh}{m_{\text{sh}}}}({\htmlData{sym=t_time}{t}}) &= \begin{cases} \dfrac{{\htmlData{sym=P_curve}{P_{\text{curve}}}}({\htmlData{sym=Sirr}{S}}_1,\dots,{\htmlData{sym=Sirr}{S}}_{\htmlData{sym=N_sub}{N}})}{{\htmlData{sym=P_curve}{P_{\text{curve}}}}({\htmlData{sym=Sirr}{S}},\dots,{\htmlData{sym=Sirr}{S}})} & \text{exact diode solve (c-Si, curve)} \\[4pt] 1 - \dfrac{{\htmlData{sym=Sirr}{S}}_{\text{beam}}}{{\htmlData{sym=Sirr}{S}}}\,\overline{{\htmlData{sym=f_worst}{f_{\text{worst}}}}} & \text{PVsyst partition (c-Si, no curve)} \\[4pt] 1 - \dfrac{{\htmlData{sym=Sirr}{S}}_{\text{beam}}}{{\htmlData{sym=Sirr}{S}}}\,{\htmlData{sym=f_mean}{f_{\text{mean}}}} & \text{area-proportional (thin-film)} \end{cases} \\ {\htmlData{sym=S_k}{S_k}} &= {\htmlData{sym=Sirr}{S}}_{\text{diff}} + {\htmlData{sym=Sirr}{S}}_{\text{gnd}} + (1 - {\htmlData{sym=f_worst_k}{f_{\text{worst},k}}})\,{\htmlData{sym=Sirr}{S}}_{\text{beam}} \end{aligned}
(4.6) Two-axis field — congruent shadows & azimuth backtracking (unison)
Δsh=d⃗−s^ d⃗⋅n^s^⋅n^,cos⁡θ=cos⁡2β cos⁡ψ+sin⁡2βψ=arg⁡min⁡ ∣ψ∣    s.t. every pair overlap=0,  ∣ψ∣≤60∘\begin{aligned} {\htmlData{sym=Delta_sh}{\Delta_{sh}}} &= {\htmlData{sym=d_vec}{\vec{d}}} - {\htmlData{sym=s_hat}{\hat{s}}}\,\frac{{\htmlData{sym=d_vec}{\vec{d}}}\cdot{\htmlData{sym=n_hat}{\hat{n}}}}{{\htmlData{sym=s_hat}{\hat{s}}}\cdot{\htmlData{sym=n_hat}{\hat{n}}}}, \qquad \cos{\htmlData{sym=theta}{\theta}} = \cos^2{\htmlData{sym=beta}{\beta}}\,\cos{\htmlData{sym=psi_bt}{\psi}} + \sin^2{\htmlData{sym=beta}{\beta}} \\ {\htmlData{sym=psi_bt}{\psi}} &= \arg\min\,|{\htmlData{sym=psi_bt}{\psi}}| \;\;\text{s.t. every pair overlap} = 0,\; |{\htmlData{sym=psi_bt}{\psi}}| \le 60^{\circ} \end{aligned}
SymbolMeaningUnitsValue / source
fvisHorizon beam visibility fraction—eq 4.1 — four 15-min substeps per hour; exactly 1 for clear hours, 0 for fully blocked; converged (within 0.02 pp of 60 substeps) — the residual is the data's own hourly sampling
hHorizon elevation profile h(φ)°API request — (compass azimuth, elevation) points, piecewise-linear with wrap-around; a profile peaking under 2° is insignificant (PVsyst threshold) and resolves to no horizon at all
fhorHorizon diffuse sky-view factor—eq 4.2 — constant per orientation (PVsyst; Quaschning & Hanitsch 1998); one factor per sweep tilt, tracker bilinear on a 7° × 15° grid; a flat horizon at elevation e on a horizontal plane gives exactly cos²e (test-pinned)
fnearNear-object diffuse silhouette factor—eq 4.4 — exact ray/rectangle mask on the same cosine-weighted dome, constant per orientation; multiplicative with fhor (conservative on overlapping silhouettes)
fworst,kWorst-strip beam shade fraction, substring k—eq 4.3 — exact corner projection + clip on the strip grid; the c-Si diode response keys on the worst strip, not the average (worst-cell-limited doctrine)
fmean,kMean beam shade fraction, substring k—eq 4.3 — area-proportional mean over each substring's strips, then module-wide to fmean; the thin-film loss path exactly (monolithic construction, no substring knock-out)
SkPer-substring irradianceW/m²eq 4.5 — diffuse + ground + (1 − fworst,k) · beam; each substring's single-diode curve is evaluated at its own Sk
mshNear-field power multiplier—eq 4.5 — ratio of shaded to unshaded curve power, so the curve level cancels and the linear path drives the energy; exactly 1 at night and for unshaded hours; scales power on both configurations
KSub-hour gate substeps—4 (15-minute) — cuts the ON/OFF sampling-phase bias from 1.39 pp to 0.19 pp of annual DNI on the 40°-wall stress case; within 0.02 pp of 60 substeps
N × CStrip grid — substrings across × cell strips along—N = bypass-diode count, C = cells in series ÷ N, from the module's electrical block; generic 3 × 20 without one (portrait layout)
footprintModule footprint for the near-field projectionmthe datasheet length × width when supplied (portrait-normalised: the short side runs across the substrings); else derived from module area at the standard 2:1 portrait aspect (1.0 × 2.0 m fallback) — echoed as module_width_m / module_height_m either way
topologyModule electrical topology for the shade solve—optional declaration, echoed as electrical_topology: standard (3 series substrings, the default) / half-cut (twin parallel pairs sharing the diodes) / back-contact (two-strategy soft-breakdown, no bypass events) / shingled (partition tier)

Deliberately not modelled (each named; conservative in direction unless stated): obstruction reflectance — far obstructions keep ground-level radiance (PVsyst-consistent); penumbra — shadows are hard-edged (a ≲ 9 cm soft edge at 10 m); multi-object overlap — the union 1 − Π(1 − fi) counts shared shade twice; the ground-reflected term keeps its full view factor; the bifacial rear's sky view is uncorrected (a rear sky-strip correction is a planned refinement — conservative at most rear share × shade loss); the tilt sweep and the §10.3 αr tables are near-field-free (a planned refinement — the optimum tilt can shift under near-field shade). For P90: shading scales the live P50 anchor, so the absolute P90/P75 bounds scale with it; the interannual CV is first-order unaffected — the horizon is static, not a year-to-year driver.

What the response reports: a shading block with a significant horizon (annual beam-loss percentage, per-configuration diffuse factors) and a near_shading block with objects (object count, the electrical tier taken, annual loss per configuration against the same request without near-field, per-configuration diffuse factors, the footprint used, and the exact tier's bypassed hours) — each absent entirely when its input block is not supplied. An array request block (drawn interactively in the site editor — envelope, walls, heights — or supplied raw over the API) additionally runs this machinery across a whole panel layout: an operating envelope is filled with equator-facing fixed-tilt panels — or, in two_axis mode, a pedestal tracker lattice tracking in unison with azimuth backtracking (rotating off-sun at low sun so no pedestal shades another; hours no rotation can clear keep their residual shading evaluated electrically, and the incidence cost is echoed separately) — each panel evaluated against the site's obstructions and its neighbours (the winter-solstice spacing rule falls out of the same geometry), with the interior-row bifacial rear model when the module is bifacial — in pedestal tracker fields the rear side follows the tracker's hourly orientation, with the spacing “behind” the rear face measured across the pedestal lattice as the field slews (treating the discrete lattice as continuous rows, a deliberately conservative simplification), and worked examples take a nominal ground reflectance of 0.20 unless the site specifies its surface — per-panel losses, array means and every assumption echoed. Sloped sites: optional per-corner ground heights fit a best-fit ground plane through the envelope, and panels (or pedestal hubs) follow it — racking stays plumb, and the exact 3-D shading machinery consumes the true height offsets, so a row dropping downhill genuinely shades less (the response echoes the fitted slope and its downhill direction; the spacing rule itself remains a flat-ground rule and the solver warns above 3° slope). Split systems: a request can carry up to four array envelopes — separate roof faces or paddocks, each with its own layout mode, spacing and corner heights — evaluated independently and reported per array; arrays are treated as disjoint faces, so panels in one never shade another (a named limitation echoed in every multi-array response). Fixed arrays can also face any compass direction (manual layouts — e.g. a real east/west roof split): the solver mode keeps its equator-facing spacing doctrine, and an off-equator choice is echoed in the assumptions. Facing direction chiefly reshapes WHEN power arrives across the day — the quantity battery sizing consumes. Shade-tolerant module classes (half-cut, back-contact, thin-film) can additionally request a density_search: a bounded tilt-and-pitch ranking (including an east–west back-to-back arrangement) that reports whether packing more panels than the no-shading rule allows would raise total output — recommendations only, never applied automatically, with the financial verdict deferred to the economics phase. Finally, an opt-in request flag (include_array_hourly) adds the array's composed hourly DC energy series to each block (8,760 kWh values through the same per-panel thermal and low-light chain the single-module yields use). By default every panel runs at its own maximum power point — the module-level-electronics (micro-inverter / optimizer) architecture. Setting panels_per_string instead wires panels into series strings: with the module's I–V block each panel's single-diode curve is summed at its own shaded irradiance, so the string's shade-coupling loss emerges from the physics hour by hour (a measured-testbed comparison of the two architectures under shade spans roughly +4% to +12% for module-level electronics — Deline et al., NREL, 2012 — and the response reports YOUR design's own delta rather than any generic figure); without the I–V block only a constant manufacturing-dispersion factor applies (defaults 2% crystalline / 1% thin-film, following PVsyst's array-mismatch documentation, 2026), and shade coupling is named as not computable. The hourly series follows the chosen architecture, always before charge-controller or inverter conversion: the series the coming battery and system layers consume. With an inverter configured, an AC series is added beside it: the PVWatts part-load efficiency model (NREL, PVWatts Version 5 Manual, 2014) anchored at the datasheet's printed weighted efficiency (the CEC weighting — an energy-averaged figure, so anchoring there reproduces the sheet's own energy claim), with output clipped at the rated AC power and any printed night-time standby consumption drawn at dark hours — the AC series can genuinely read negative overnight, which is exactly what battery arithmetic needs to see.

Tracker backtracking, in brief (eq 4.6): because every pedestal in the field holds the same orientation each hour, a neighbour’s shadow on a panel is always a same-size, shifted copy of the panel itself (Fernández-Ahumada et al. 2019; Jensen et al. 2022) — so “does anything shade anything?” reduces to a closed-form overlap test over every pair offset (Fig. 4.3), and backtracking is the SMALLEST shared rotation of the field’s pointing direction away from the sun that makes every overlap vanish (Lorenzo, Narvarte & Muñoz 2011, generalised; Fig. 4.4). Smaller rotation always means smaller pointing loss (Fig. 4.5), so minimal rotation is provably the cheapest shade-free choice; hours where no allowed rotation clears the field keep the best available rotation and their remaining shade goes through the module’s real electrical response — never dropped. Why rotate in azimuth rather than tilt back? Tilting the panels back toward horizontal also clears shadows, but at a larger pointing cost for the same relief. The field-scale study behind this design found azimuth backtracking roughly energy-neutral — the pointing loss about equals the shade loss it avoids — and it still wins decisively, because clearing the shade also eliminates hot-spots, the concentrated cell heating that partial shade drives through real modules (Lorenzo, Narvarte & Muñoz 2011). A secondary tilt-back refinement exists in that same work and remains a recognised future option here. The whole construction uses compass-free vector geometry: no latitude or hemisphere term appears anywhere, so the same code serves Perth, Nairobi and Madrid (Marion & Dobos 2013; Anderson & Mikofski 2020 lineage). Two refinements are built into every two-axis simulation, not offered as options. The field’s dodges are DIFFUSE-AWARE: each engaged hour is re-aimed at the brightest overall shade-free sky rather than the smallest rotation — searching tilt and azimuth together, both faces when bifacial (the depth note below explains when and why this pays). And the field is OBSTRUCTION-AWARE: a pedestal whose sun is hidden behind an obstruction stops chasing what it cannot see, turns to face the brightest patch of open sky, and — crucially — leaves the backtracking solve, so the pedestals that can still see the sun never dodge shadows their blocked neighbour was going to receive anyway. The stowed pedestal’s orientation is chosen under a strict guard: it may never cast shade onto a tracking neighbour. On an open site nothing is ever blocked and the field simply moves as one. Why denser can beat “no shading”: the classic spacing rule buys zero shading by spending land. Module classes that degrade gracefully under partial shade (half-cut, back-contact, thin-film) can trade a few percent of per-panel loss for substantially more panels in the same envelope — the density search quantifies exactly this trade, and an east–west back-to-back arrangement (rows facing away from each other, sharing their high edge) is often the land-efficiency winner despite the counter-intuitive tilt (up to +1.7% over full avoidance is documented for half-cut modules even at row scale; Dobos 2022, PVPMC). Terrain: the far-horizon block can be prefilled from elevation data (PVGIS terrain profile — free, terrain only, never trees or buildings; the prefill says so and stays editable), and per-corner ground heights already tilt the array’s ground plane (above). Those corner heights can themselves be prefilled from Google’s aerial surface model where coverage exists (0.1–0.25 m imagery; requested at the lowest quality floor so coverage gaps are honest misses): the values are SURFACE heights — buildings and trees count where they stand — taken relative to the pin, and the capture date is shown beside the prefill because aerial surveys can be years old (source: Google Solar API; includes solar data from Google). Nothing is stored server-side — only the heights you keep in the fields. Automatic obstruction extraction remains planned premium-tier work, documented here as it lands, never before.

Animation — a worked example — Perth on the winter solstice — drawn in the sun's own view, rendered by the same solver every simulation runs. Looking along the beam, a panel and its shadow are the same shape, so panels overlap in this view exactly when one shades another; red patches are beam-blocked area (eq 4.6). At first light no allowed rotation can clear the field — the rotation caps at 60° and the small red overlaps that remain are carried through the electrical model. As the sun climbs the field backtracks: the silhouettes glide apart until they just touch — the smallest clearing rotation, visible as tangency — and by mid-morning they separate fully (green outlines). The evening mirrors the morning. Play it, or step through frame by frame.
pedestal 1pedestal 2panel 1panel 2 — same tiltŝshadow — a shifted copy of panel 1dn̂Δsh = d − ŝ·(d·n̂)/(ŝ·n̂)the offset, slid along the light
Fig. 4.3 — In a unison tracker field every aperture is parallel — and on level ground every hub is level, so the centre offset d is a horizontal vector. A neighbour's shadow is a congruent, translated copy of the panel itself: d projects along the sun direction ŝ into the aperture plane (drawn perpendicular to the beam — sun-normal), shifting by Δsh (eq 4.6). Whether anything shades anything is then a closed-form overlap test between two same-size rectangles — per cell-strip, a 1-D interval overlap — with no polygon clipping (Fernández-Ahumada et al. 2019; Jensen et al. 2022). Original figure.
low suntracking sun-normal — overlapsψ = 0 · shadows hit the next rowψbacktracked — every overlap goneone shared ψ · smallest |ψ| · |ψ| ≤ 60°
Fig. 4.4 — Backtracking in plan view. Left: at low sun the sun-normal field's shadows (dashed, congruent copies) land on the down-sun neighbours — the hatched overlaps. Right: the whole field rotates its pointing azimuth by one shared ψ — found by a symmetric outward scan then bisection, smallest |ψ| first, capped at ±60° — until every pair overlap vanishes (eq 4.6). Note the shadows' direction never changes — the sun fixes it; the rotation only changes the apertures' footprint, sliding the copies past the neighbours. Hours no allowed ψ can clear keep the overlap-minimising rotation and the residual shade is evaluated electrically, never dropped. Original figure.
-60°-40°-20°0°20°40°60°0.40.60.81.0backtracking rotation ψcos θβ = 10° — low sunβ = 35°β = 60° — high sunψ* — smallest clearing rotationeven in ψ — both directions cost the same
Fig. 4.5 — What a rotation costs: cos θ = cos²β·cos ψ + sin²β (eq 4.6), the fraction of sun-normal beam the aperture keeps after rotating ψ away. The curve is even in ψ — either direction costs the same, so the search only cares about |ψ| — and flat near ψ = 0, so the smallest clearing rotation is provably the cheapest shade-free choice. At high sun (β = 60°) the curve is nearly flat everywhere: backtracking is almost free exactly when shadows are shortest and it is least needed. Original figure, drawn from the model's own closed form.

One field angle, or one angle per pedestal? A fully greedy alternative exists: free each pedestal to rotate only as far as its own sun-side neighbours force it, the most sunward unit never rotating at all. The animation below shows it — and shows why this model rejected it. A sun-tracking sunward unit casts its widest shadow, so greedy freedom manufactures shade in tight interiors and loses to the one-motion field on open sites (measured here at several percent on a winter day). What greedy got right is kept: its obstruction awareness — a pedestal whose sun is hidden stops chasing what it cannot see and stows for diffuse light. The production algorithm is the synthesis: the field moves as one, and ONLY a blocked pedestal decouples — under a guard that forbids it from shading the tracking field, and with its geometry removed from the backtracking solve so it cannot force needless dodging either.

Animation — the same field with a wall to the east, in the same sun's-eye view — which now shows why pedestals stow: the wall's slate silhouette is the part of the sky it steals, and a pedestal sitting inside it cannot see the sun at all (its red fill is beam blocked by the wall). Under per-unit backtracking each pedestal is solved sun-side first — the most sunward unit never rotates, the rest rotate only as far as their own sun-side neighbours force them (amber) — and a blocked pedestal stops chasing a sun it cannot see, stowing to the orientation that collects the most diffuse light from the open sky (blue). Freedom is not free: a sun-tracking sunward unit casts its widest shadow, so the greedy field manufactures shade the one-motion field avoids — which is why the production algorithm keeps only greedy's obstruction stow, grafted onto the one-motion field. Play it, or step through frame by frame.

Two behaviours worth watching for. The stow angle: a stowed pedestal is not idle — with its sun hidden it cannot harvest beam at all, so the best it can do is collect diffuse light, and it turns to face the brightest patch of open sky. Near a large obstruction that patch is not overhead: the wall eats a slab of the sky dome, so the diffuse-optimal orientation leans away from it — at angles that can look odd until you remember what the panel is actually optimising. The evening scramble: toward sunset the roles reverse. The western pedestals become the sun side of the field and track freely — and a sun-tracking pedestal casts its widest shadow — while the now-leeward eastern pedestals swing hard between clearing rotations as shadows lengthen faster than rotation can clear them; hours where nothing clears keep the best available rotation (red) and the remaining shade is evaluated electrically. That scramble is the greedy trade-off made visible — and exactly why the production field moves as one, decoupling only pedestals whose sun an obstruction has already taken. The next animation replays the identical scene under the production algorithm: the same dawn that scatters the greedy field into reds and swings is handled with a smaller, calmer dodge — and a row of stowed pedestals quietly harvesting sky.

Animation — the same walled scene, solved by the production algorithm. At dawn the pedestals the wall blinds stow (blue) — normals aimed at the open sky, silhouettes held clear of the tracking field by the admissibility guard — and, because their geometry leaves the backtracking solve, the heads-up display shows the field dodging by less than a full-field solver would demand: blocked pedestals force no needless backtracking. Violet marks hours the dodge is re-aimed at the brightest shade-free sky (the diffuse-aware search at work inside the core). And where the greedy animation ends its day in a scramble, this field ends it moving as one — the wall is down-sun by evening, so nothing is blocked and nothing needs freedom. Play it, or step through frame by frame.

How this relates to published and commercial practice. Per-row backtracking is established in the single-axis literature, where the driver is terrain: slope-aware backtracking derives each row’s rotation from cross-axis slope (Anderson & Mikofski 2020, NREL), closed-form shaded fractions extend it to rolling ground (Sandia 2024), and a linear-programming formulation computes per-row rotations for hundreds of trackers, including deliberate “fractional backtracking” under an allowed-shade budget (2025). Commercially, smart tracker control is validated at the 1–2% level in independent field assessments (Nextracker TrueCapture, 2022–24): row-to-row backtracking, a diffuse mode that flattens under bright overcast skies, and Split Boost — which deliberately admits up to half-module shade on half-cut modules, trading bounded electrical loss for steeper tracking. Those systems are machine-learned and proprietary. The algorithm here is the transparent counterpart for two-axis fields, and differs in what drives the decoupling: not terrain but obstructions — pedestals are decoupled exactly when a site obstruction hides their sun, under an exact geometric guard, with the diffuse-aware re-aim built on the published maximum-irradiance formulation (Fernández-Ahumada et al. 2019/2020) rather than learned behaviour. Every step is stated in this methodology and reproducible from it. The admitted-shade trade that Split Boost and fractional backtracking exploit is modelled here too, and gated by the architecture that makes it safe: on half-cut modules — which this model resolves as twin substrings (§7) — the backtracking condition relaxes from “no shade” to “upper halves protected”, with the bounded lower-half shade priced exactly by the electrical tier. Building it produced a finding worth stating: in two-axis fields the sweep of a neighbour's shadow is fast, so the window between silhouettes touching and a half-buried module is narrow — steeper tracking on otherwise-clearable hours essentially never pays, and the trade's real two-axis value is rescuing the hours no rotation can fully clear: bounded, upper-protected shade instead of unbounded mixed shade. Split Boost's cleared-hour gains belong to the dense single-axis rows it was built for.

Is sun-normal even the right target? A fair question at very low sun, when much of what remains is diffuse light gathered near the horizon. The pointing target here is deliberately beam-first — and less naive than it looks: the sky's circumsolar share travels with the beam in this model's irradiance treatment (§5), so sun-normal already collects the brightening around the sun; and the shade-free condition admits two equal-beam rotations (flatter or steeper), with the flatter chosen partly because downwelling diffuse favours it. Orienting to maximise the full direct-plus-diffuse mix instead — maximum-irradiance backtracking — has been shown in simulation to add about +1.3% annually at one mid-latitude site, varying from a December penalty to an August gain (Fernández-Ahumada et al. 2020), with published upper bounds near 3% at high-diffuse sites (Solar RRL 2024). This model builds exactly that refinement into every two-axis simulation: each engaged hour is re-aimed at the shade-free orientation collecting the most total light, and the tracking telemetry beside the results reports how often it mattered for your site. The search spans both axes — a subtlety worth stating: at sun-normal tilt, rotating in azimuth alone leaves the diffuse view unchanged, so the diffuse gain lives in tilting back — and on bifacial modules the objective prices both faces, because a diffuse-led orientation must also answer for the rear, where lying flatter costs what the front gains (NREL field results, 2025). One boundary is deliberate: for standard modules the search only considers orientations that are exactly shade-free, never a little shade traded for a wider sky view (half-cut modules earn a bounded exception — the admitted-shade passage above). Collected irradiance is an honest yardstick between two unshaded orientations, but partial shade costs more than the area it covers at every electrical level: within a module, a sliver across one cell strip sacrifices its whole bypass substring (the worst-strip rule above), and modules wired in a series string are dragged toward the shaded module's current unless each carries its own electronics — a contrast priced explicitly in §18. A partially shaded candidate that looks brighter on paper would therefore usually deliver less. The shade-free constraint is how the published method is formulated (Fernández-Ahumada et al. 2019), and it is why the search re-aims and tilts back but never lets the silhouettes overlap.

Animation — the same field and the same day, under diffuse-aware backtracking. Whenever the standard solver would dodge, the search instead spans tilt and azimuth together for the shade-free orientation collecting the most total light. Violet marks the diffuse-aware choice, and the geometry explains itself: tilting back squashes each silhouette vertically — that is the beam sacrifice, paid in exchange for a flatter aperture that sees more of the bright sky — and the normal arrows lean up and away from you (the sun). Watch the first and last hours especially: rotations that cap out red in the standard animation are cleared outright here, because shade that azimuth-only rotation cannot escape often yields to the second axis. Where a gain figure is shown it is the solver's own accounting for that hour; hours with no shade-free baseline report the clearance instead of a delta. Play it, or step through frame by frame.

5. Two-Axis Tracking — Incident Irradiance▾

A panel that always faces the sun has an incidence angle of ~0°, so beam irradiance equals DNI with no cosine penalty and no beam reflection (IAM) penalty. Sky-diffuse light transposes by the Hay–Davies model: the circumsolar share Id·Ai·Rb joins the beam (it arrives from the sun's direction — see the concept box below), while the isotropic remainder and the ground-reflected component use the standard view factors at the instantaneous tilt Σ = 90° − β, each multiplied by its integrated Martín–Ruiz angular-loss factor Fd(Σ) / Fa(Σ) — sky and ground light arrive obliquely, so glass reflection still costs them a few percent even on a tracker. The closed forms are given in full below (eq 5.2); they are the same forms as §6's eq 6.3, evaluated here at the tracker's hourly tilt — no single incidence angle exists for diffuse light, so the loss is an integral over the view. All components are zero when the sun is below the horizon (β ≤ 0).

Concept — chasing the sun kills the cosine loss

Hold a flashlight square to a wall and it makes a small, bright circle; tilt it and the same light smears into a long, dim ellipse. That is the cosine loss: a surface intercepts only cos θ of a parallel beam, where θ is the angle between the beam and the surface normal (Fig. 5.1). At θ = 60° a panel captures exactly half the beam — before a single photon is reflected or absorbed. A two-axis tracker's entire job is to hold θ ≈ 0 from sunrise to sunset, so the beam term is simply the full DNI, every hour. This — not more hours of daylight — is where most of the tracking gain comes from.

same parallel beamθ = 0 — full captureθ = 60°same panel at θ = 60° — half the beam (cos 60° = 0.5)
Fig. 5.1 — The cosine loss. Left: a panel square to a parallel beam intercepts its full width. Right: the same panel turned 60° away from the beam intercepts only cos 60° = half of it — like a flashlight held at a slant, the same light spreads over more surface, so each square metre receives less. A two-axis tracker holds θ ≈ 0 all day, taking the full beam every hour; every fixed panel pays cos θ for every hour the sun is off its normal. Original figure.

Concept — what the panel sees — view factors

Diffuse and ground-reflected light have no direction to chase, so they follow a different rule: how much sky and how much ground fills the panel's field of view. Lie flat on your back and you see the entire sky and no ground; stand a panel bolt upright and it sees half of each. The isotropic view factors (1 + cos Σ)/2 and (1 − cos Σ)/2 interpolate between those extremes. A tracker following a 34.6° winter-noon sun stands at Σ = 55.4°, seeing 78% sky and 22% ground — so tracking trades a little diffuse capture for the full beam. With DNI usually several times the diffuse, that trade wins whenever the sun is out.

Concept — the sky is brighter around the sun — the Hay–Davies split

Real skies are not equally bright everywhere: a bright halo surrounds the sun (circumsolar light — forward-scattered by haze and thin cloud), and validation against instrumented tilted sensors shows that ignoring it under-reads panels facing the sun (Loutzenhiser et al. 2007). This model uses the Hay–Davies sky (Hay & Davies 1980): each hour, a fraction Ai = DNI/E0 of the diffuse light — the anisotropy index, the sky's own measured clearness — is treated as arriving from the sun's direction, and the remainder as the uniform dome the view factors above describe. The circumsolar share behaves like beam light everywhere downstream: it carries the beam reflection loss, disappears when terrain hides the sun, and is blocked by the same obstructions that shade the beam. Under overcast (DNI = 0) the split vanishes and the sky is purely uniform. Ai is a physical ratio with no fitted coefficients — the reason this model was chosen over the Perez model, whose empirically fitted sky bins and modelled horizon band sit poorly beside this tool's measured horizon profiles. Benchmarks rank Hay–Davies among the most accurate transposition models when driven by standard weather-file inputs (Yang 2016, Solar Energy 136). Residual differences vs Perez-based software (PVsyst, SAM) are ~0–3%, in either direction (verified against pvlib and NREL SAM at six climates, 2026).

(5.1) Tracker faces the sun (θ ≈ 0)
Σ=90∘−β [∘]{\htmlData{sym=Sigma}{\Sigma}} = 90^\circ - {\htmlData{sym=beta}{\beta}}\ [^\circ]
(5.2) Hay–Davies anisotropy index
Ai=DNI/E0,Rb=1/max⁡(sin⁡β, cos⁡89∘){\htmlData{sym=A_i}{A_i}} = {\htmlData{sym=DNI}{\text{DNI}}}/{\htmlData{sym=E_0}{E_0}}, \qquad {\htmlData{sym=R_b}{R_b}} = 1/\max(\sin{\htmlData{sym=beta}{\beta}},\ \cos 89^\circ)
(5.3) Integrated diffuse/ground factors (Martín–Ruiz 2001) — at the hourly tilt
Fd=1−e−(c1+c2Td) Td/αrFa=1−e−(c1+c2Ta) Ta/αr{\htmlData{sym=F_d}{F_d}} = 1 - {\htmlData{sym=e_euler}{e}}^{-({\htmlData{sym=c_1}{c_1}} + {\htmlData{sym=c_2}{c_2}} {\htmlData{sym=T_dterm}{T_d}})\,{\htmlData{sym=T_dterm}{T_d}}/{\htmlData{sym=alpha_r}{\alpha_r}}} \qquad {\htmlData{sym=F_a}{F_a}} = 1 - {\htmlData{sym=e_euler}{e}}^{-({\htmlData{sym=c_1}{c_1}} + {\htmlData{sym=c_2}{c_2}} {\htmlData{sym=T_aterm}{T_a}})\,{\htmlData{sym=T_aterm}{T_a}}/{\htmlData{sym=alpha_r}{\alpha_r}}}
(5.4) Beam + circumsolar — no cosine loss, beam IAM = 1
Ibeam=DNI+Id Ai Rb [W/m2]{\htmlData{sym=I_beam}{I_{\text{beam}}}} = {\htmlData{sym=DNI}{\text{DNI}}} + {\htmlData{sym=I_d}{I_d}}\,{\htmlData{sym=A_i}{A_i}}\,{\htmlData{sym=R_b}{R_b}}\ [\text{W/m}^2]
(5.5) Diffuse remainder — isotropic background × F_d
Idiffuse=Id (1−Ai) 1+cos⁡Σ2 Fd(Σ) [W/m2]{\htmlData{sym=I_diffuse}{I_{\text{diffuse}}}} = {\htmlData{sym=I_d}{I_d}}\,(1 - {\htmlData{sym=A_i}{A_i}})\,\dfrac{1 + \cos{\htmlData{sym=Sigma}{\Sigma}}}{2}\,{\htmlData{sym=F_d}{F_d}}({\htmlData{sym=Sigma}{\Sigma}})\ [\text{W/m}^2]
(5.6) Ground-reflected × F_a
Iground=Ir 1−cos⁡Σ2 Fa(Σ) [W/m2]{\htmlData{sym=I_ground}{I_{\text{ground}}}} = {\htmlData{sym=I_r}{I_r}}\,\dfrac{1 - \cos{\htmlData{sym=Sigma}{\Sigma}}}{2}\,{\htmlData{sym=F_a}{F_a}}({\htmlData{sym=Sigma}{\Sigma}})\ [\text{W/m}^2]
(5.7) Plane-of-array total
Itotal=Ibeam+Idiffuse+Iground [W/m2]{\htmlData{sym=I_total}{I_{\text{total}}}} = {\htmlData{sym=I_beam}{I_{\text{beam}}}} + {\htmlData{sym=I_diffuse}{I_{\text{diffuse}}}} + {\htmlData{sym=I_ground}{I_{\text{ground}}}}\ [\text{W/m}^2]
SymbolMeaningUnitsValue / source
ΣInstantaneous collector tilt (faces the sun)°90° − β, recomputed every hour
AiAnisotropy index — the circumsolar fraction of diffuse light—DNI / E0 (extraterrestrial normal); 0 overcast, ~0.8 clearest; from the horizon-gated DNI
RbBeam projection ratio onto the sun-tracking plane—1 / max(sin β, cos 89°) — the floor stops sunrise spikes
IbeamBeam-like light on the panel (beam + circumsolar)W/m²DNI + Id × Ai × Rb (θ ≈ 0, no cosine loss)
IdiffuseIsotropic-remainder diffuse on the panelW/m²Id × (1 − Ai) × (1 + cos Σ)/2 × Fd(Σ) — remainder × sky view factor × integrated angular loss
IgroundGround-reflected component on the panelW/m²Ir × (1 − cos Σ)/2 × Fa(Σ) — ground view factor × integrated angular loss
Fd, FaIntegrated Martín–Ruiz angular-loss factors — full forms in eq 5.2 (≡ §6 eq 6.3)—evaluated at the tracker's hourly tilt; same formulas as the fixed panel — like-for-like
ItotalTotal plane-of-array (POA) irradianceW/m²sum of the three components
6. Fixed-Tilt Reference System▾

A stationary panel at tilt Σ = |latitude|°, facing the equator (due north in the southern hemisphere, due south in the northern). The incidence angle θ between the sun and the panel normal varies continuously (thesis eq 8); the Martín–Ruiz incidence angle modifier (IAM) then accounts for reflection losses at oblique angles (thesis eq 20 — stated there as the angular loss; IAM is its complement). Beam is zeroed when θ ≥ 90° or the sun is below the horizon.

Concept — the panel's-eye view — the incidence angle θ

Everything the fixed panel loses relative to the tracker is captured by one angle: θ, between the incoming beam and the panel's normal (Fig. 6.1). Computing it is a change of perspective — the sun's sky position (β, φs, §1) is re-measured against the panel's own orientation (tilt Σ, facing direction φp). Thesis eq 8 does this in one spherical-trigonometry step, and only the difference of the two bearings (φs − φp) enters it (Fig. 6.2): swing the whole scene around the compass and nothing changes. θ then drives two separate penalties — the cos θ footprint spreading (Fig. 5.1) and the reflection loss below.

fixed panelΣ = 32°panel normalhorizontalβ = 34.6°θ
Fig. 6.1 — Side view of the fixed-tilt geometry, drawn to scale for Perth at noon on 21 June: panel tilt Σ = 32° (= |L|), sun elevation β = 34.6°, so the beam arrives θ = 23.4° off the panel normal. θ is the single angle the whole fixed-tilt calculation turns on: the beam is scaled by cos θ (footprint spreading, Fig 3.1) and by IAM(θ) (reflection, Fig 4.3). Original figure; geometry after thesis eq 8 / Masters (2013).
NWEpanel normalφₚφₛviewed from above — this model: φₚ = 0° (equator-facing, southern hemisphere)
Fig. 6.2 — The same geometry seen from above. Both azimuths are compass bearings from north: φₛ locates the sun, φₚ is the direction the panel faces. Only their difference (φₛ − φₚ) enters the incidence equation — swing the sun around the panel and only the relative bearing matters. Drawn general; in this model the fixed panel faces the equator exactly (φₚ = 0° in the southern hemisphere, 180° in the northern), so the difference reduces to φₛ itself. Original figure.

Concept — glass turns into a mirror — the IAM

Look at a lake at noon and you see the bottom; look at it near sunset and it turns to glare — at grazing angles, any smooth dielectric surface reflects rather than transmits. Panel glass is no exception: light arriving far off the normal increasingly bounces off before reaching the cell. The Martín–Ruiz IAM (Fig. 6.3) quantifies the transmitted fraction: essentially 1 out to θ ≈ 55°, still 95% at 60°, then a cliff — 87% at 70°, 64% at 80°. Its single parameter αr describes how "rough" the optical surface is: 0.17 for clean commercial glass, rising to 0.27 for heavily soiled glass (§10.3), because a dust layer scatters extra light at every angle. The two-axis tracker sits permanently at the flat top of this curve — a second, independent win over the fixed panel.

0.250.50.7510°15°30°45°60°75°90°angle of incidence θIAM — transmitted fraction of the beamαᵣ = 0.17 — clean0.19 — 10% soiled0.20 — 20% soiled0.27 — 30% soiled● the knee — θ = 60°: a clean panel still transmits 95%
Fig. 6.3 — The Martín–Ruiz incidence angle modifier — the fraction of the beam that gets through the glass instead of glancing off — computed from the model's own formula at its four soiling levels. The curve is remarkably forgiving until ~60° (a clean panel still transmits 95%), then falls off a cliff: 87% at 70°, 64% at 80°. Soiling drags the whole curve down (αᵣ 0.17 → 0.27), because dusty glass scatters extra light at every angle. Two-axis tracking lives permanently at the flat top (θ ≈ 0). Original figure; formula per Martín & Ruiz (2005), thesis eq 20.
(6.1) Incidence angle (thesis 8)
cos⁡θ=cos⁡β cos⁡(φs−φp) sin⁡Σ+sin⁡β cos⁡Σ\cos{\htmlData{sym=theta}{\theta}} = \cos{\htmlData{sym=beta}{\beta}}\,\cos({\htmlData{sym=varphi_s}{\varphi_s}} - {\htmlData{sym=varphi_p}{\varphi_p}})\,\sin{\htmlData{sym=Sigma}{\Sigma}} + \sin{\htmlData{sym=beta}{\beta}}\,\cos{\htmlData{sym=Sigma}{\Sigma}}
(6.2) Martín–Ruiz beam IAM (thesis 20)
IAM(θ)=1−exp⁡(−cos⁡θ/αr)1−exp⁡(−1/αr){\htmlData{sym=IAM}{\text{IAM}}}({\htmlData{sym=theta}{\theta}}) = \dfrac{1 - \exp(-\cos{\htmlData{sym=theta}{\theta}} / {\htmlData{sym=alpha_r}{\alpha_r}})}{1 - \exp(-1/{\htmlData{sym=alpha_r}{\alpha_r}})}
(6.3) Integrated diffuse/ground factors (Martín–Ruiz 2001 · IEC 61853-3)
Fd=1−e−(c1+c2Td) Td/αrFa=1−e−(c1+c2Ta) Ta/αr{\htmlData{sym=F_d}{F_d}} = 1 - {\htmlData{sym=e_euler}{e}}^{-({\htmlData{sym=c_1}{c_1}} + {\htmlData{sym=c_2}{c_2}} {\htmlData{sym=T_dterm}{T_d}})\,{\htmlData{sym=T_dterm}{T_d}}/{\htmlData{sym=alpha_r}{\alpha_r}}} \qquad {\htmlData{sym=F_a}{F_a}} = 1 - {\htmlData{sym=e_euler}{e}}^{-({\htmlData{sym=c_1}{c_1}} + {\htmlData{sym=c_2}{c_2}} {\htmlData{sym=T_aterm}{T_a}})\,{\htmlData{sym=T_aterm}{T_a}}/{\htmlData{sym=alpha_r}{\alpha_r}}}
(6.4) Hay–Davies anisotropy index
Ai=DNI/E0,Rb=max⁡(cos⁡θ,0)/max⁡(sin⁡β, cos⁡89∘){\htmlData{sym=A_i}{A_i}} = {\htmlData{sym=DNI}{\text{DNI}}}/{\htmlData{sym=E_0}{E_0}}, \qquad {\htmlData{sym=R_b}{R_b}} = \max(\cos{\htmlData{sym=theta}{\theta}}, 0)/\max(\sin{\htmlData{sym=beta}{\beta}},\ \cos 89^\circ)
(6.5) Beam + circumsolar — single-angle IAM on sun-direction light
Ibeam=(DNI cos⁡θ+Id Ai Rb) IAM(θ) [W/m2]{\htmlData{sym=I_beam}{I_{\text{beam}}}} = ({\htmlData{sym=DNI}{\text{DNI}}}\,\cos{\htmlData{sym=theta}{\theta}} + {\htmlData{sym=I_d}{I_d}}\,{\htmlData{sym=A_i}{A_i}}\,{\htmlData{sym=R_b}{R_b}})\,{\htmlData{sym=IAM}{\text{IAM}}}({\htmlData{sym=theta}{\theta}})\ [\text{W/m}^2]
(6.6) Diffuse remainder — isotropic background × F_d
Idiffuse=Id (1−Ai) 1+cos⁡Σ2 Fd(Σ) [W/m2]{\htmlData{sym=I_diffuse}{I_{\text{diffuse}}}} = {\htmlData{sym=I_d}{I_d}}\,(1 - {\htmlData{sym=A_i}{A_i}})\,\dfrac{1 + \cos{\htmlData{sym=Sigma}{\Sigma}}}{2}\,{\htmlData{sym=F_d}{F_d}}({\htmlData{sym=Sigma}{\Sigma}})\ [\text{W/m}^2]
(6.7) Ground-reflected × F_a
Iground=Ir 1−cos⁡Σ2 Fa(Σ) [W/m2]{\htmlData{sym=I_ground}{I_{\text{ground}}}} = {\htmlData{sym=I_r}{I_r}}\,\dfrac{1 - \cos{\htmlData{sym=Sigma}{\Sigma}}}{2}\,{\htmlData{sym=F_a}{F_a}}({\htmlData{sym=Sigma}{\Sigma}})\ [\text{W/m}^2]
(6.8) Plane-of-array total
Itotal=Ibeam+Idiffuse+Iground [W/m2]{\htmlData{sym=I_total}{I_{\text{total}}}} = {\htmlData{sym=I_beam}{I_{\text{beam}}}} + {\htmlData{sym=I_diffuse}{I_{\text{diffuse}}}} + {\htmlData{sym=I_ground}{I_{\text{ground}}}}\ [\text{W/m}^2]
SymbolMeaningUnitsValue / source
θAngle of incidence between sun and panel normal°thesis eq 8; at Σ = 0 reduces to cos θ = sin β (thesis eq 7) — the sanity check
φpPanel azimuth°0° (due north) if L < 0, else 180° (due south)
ΣPanel tilt°|L| by default; §7 finds the true optima
αrMartín–Ruiz angular-loss coefficient—clean anchor 0.17 (0.14 AR-coated glass, or a measured module value) → 0.19 / 0.20 / 0.27 with rising soiling (Martín & Ruiz 2005)
IAMIncidence angle modifier — transmitted fraction of beam—1 at θ = 0; falls steeply above θ ≈ 55°
Fd, FaIntegrated diffuse / ground angular-loss factors—closed-form integrals over the sky dome / ground band (Martín & Ruiz 2001; c2 per IEC 61853-3); both share αr, so soiling degrades all three components
Td, TaGeometry terms of the F integrals—functions of tilt only (radians); complementary — F_d(Σ) = F_a(180° − Σ)
Ai, RbHay–Davies circumsolar terms (§5 concept)—the circumsolar share Id·Ai·Rb joins the beam inside the IAM bracket — here Rb = max(cos θ, 0)/max(sin β, cos 89°)

Worked example — Perth fixed-tilt (Σ = 32°, panel faces due north)

Noon, 21 June: the sun is due north at β = 34.6°, so θ = (90° − 34.6°) − 32° = 23.4° and IAM = 0.998 — at midday the fixed panel is nearly as well aimed as the tracker. 15:00 solar, 21 June: the sun has moved west (φs = +43.6°) and dropped to β = 19.8°; now θ = 49.6° — cos θ = 0.65 already forfeits a third of the beam, while IAM = 0.981 barely notices. The order matters: through mid-range angles the cosine loss does the damage, and reflection only joins in near grazing (θ = 80°: cos θ = 0.17, IAM = 0.64). (All values computed with the model's own functions.)

The single-angle IAM(θ) is applied to the beam-like light only (the true beam plus the Hay–Davies circumsolar share, which arrives from the same direction — §5) — the remaining diffuse and ground-reflected light are not collimated, so there is no single θ to plug in. That does not make them loss-free: every element of the sky dome and ground band arrives at its own oblique angle, and oblique rays reflect more. Martín & Ruiz integrated the angular loss over the isotropic radiance distributions in closed form — the factors Fd(Σ) and Fa(Σ) of eq (6.3), always below 1 (clean glass at Σ = 30°: Fd ≈ 0.96, Fa ≈ 0.80 — the ground band arrives near-grazing, so it loses most). Both configurations apply the same factors at their own tilts, so the comparison stays like-for-like; the total effect is roughly 1–2% of annual yield, more where diffuse dominates.

When soiling is present, αr is elevated: dust on glass increases angular reflectance independently of the light it blocks outright — and since the F factors share the beam's αr, soiling now degrades all three components consistently. The clean anchor is the module's own αr (0.17 generic; 0.14 for anti-reflective-coated glass, or a measured datasheet value). The simulation precomputes monthly fixed-tilt yields at all four αr breakpoints for each tilt configuration and interpolates at the effective value — the combined-loss formula is given in §10.3.

7. Tilt Optimisation▾

A 0–90° sweep in 1° steps — 91 full 8,760-hour fixed-tilt simulations — finds the tilt that maximises three objectives independently: total annual yield, winter yield, and summer yield. Winter is the coldest quarter (Dec–Feb in the northern hemisphere, Jun–Aug in the southern); summer is the reverse. Winter-optimal tilts are steeper; summer-optimal are shallower.

Bifacial modules are optimised as bifacial: each candidate tilt's yield includes its rear side (§17), so the optimum shifts a few degrees steeper over bright ground — steeper tilts show the rear more of it. The rear plane-of-array irradiance is computed on a 5° tilt grid and linearly interpolated to each 1° candidate (rear irradiance varies smoothly with tilt; the front keeps full 1° resolution).

Concept — one tilt cannot face every season

The classic rule of thumb — tilt equals latitude — has a clean geometric meaning: it points the panel exactly at the equinox noon sun, and since declination swings ±23.45° symmetrically about zero, the two solstice noons land an equal 23.4° either side of the normal (Fig. 7.1 — drawn to the model's Perth numbers). Any fixed tilt is a season-weighted compromise: steepen it and winter improves at summer's expense; flatten it and the reverse. The sweep replaces the rule of thumb with a search — 91 full-year simulations, one per degree — weighting each hour by the energy actually on offer in the measured weather, which is why the optimum lands near |L| but rarely on it: cloud seasonality, and the thermal derating baked into the objective, both shift it.

Σ = |L| = 32°panel normal21 Jun noon · β = 34.6°equinox noon · β = 58°21 Dec noon · β = 81.4°θ = 23.4°θ = 23.4°
Fig. 7.1 — Three noons, one panel (Perth, Σ = |L| = 32°). The latitude-tilt rule points the panel exactly at the equinox noon sun (θ = 0); because declination swings ±23.45° around zero, the two solstice noons land symmetrically 23.4° either side of the normal — a built-in compromise no fixed tilt can escape. The §7 sweep replaces the rule of thumb with a search: it weights every hour by the energy actually available (and by cell temperature), which is why the optimum shifts a few degrees with climate. Original figure.
(7.1) 0–90° sweep · 91 fixed-tilt sims
Σ⋆=arg⁡max⁡[ ∑tP(Σ,t) ]  over Σ∈[0,90∘]{\htmlData{sym=Sigma}{\Sigma}}^\star = \arg\max\big[\ \textstyle\sum_{\htmlData{sym=t_time}{t}} {\htmlData{sym=Ppow}{P}}({\htmlData{sym=Sigma}{\Sigma}}, {\htmlData{sym=t_time}{t}})\ \big]\ \ \text{over}\ {\htmlData{sym=Sigma}{\Sigma}} \in [0, 90^\circ]
SymbolMeaningUnitsValue / source
Σ*Optimal tilt for an objective°argmax over integer tilts 0–90°
P(Σ, t)Hourly PV power at tilt Σ (from §9)W/m²full simulation per tilt — irradiance → cell temperature → power
ΣtSum over the objective's hours—all hours (annual), coldest quarter (winter), or sunniest quarter (summer)

Because the sweep maximises power output — with thermal derating applied — rather than raw irradiance, the returned optimum is the power-optimal tilt. In hot climates it sits a few degrees steeper than the irradiance-optimal tilt, because a more vertical panel runs slightly cooler. The sweep also produces, per configuration: monthly yields (consumed by the snow model, §11), αr yield tables (consumed by the soiling model, §10), and 25-year degradation profiles (§12) for all five system configurations.

8. Cell Temperature — Faiman (2008) Thermal Model▾

Why not NOCT? The industry-standard NOCT (nominal operating cell temperature) model assumes a fixed 1 m/s wind speed. A windy site (5 m/s) actually runs ~8–10 °C cooler, yielding ~4% more output. Faiman is wind-aware, is the IEC 61853-2 standard model, and is what PVGIS itself uses internally — so the thermal model is methodologically consistent with the irradiance data source.

Concept — a heat budget in one fraction

A panel in the sun is a bathtub with the tap running: sunlight pours heat in at S watts per square metre — S is the same plane-of-array irradiance called Itotal in §5/§6; this section keeps Faiman's (2008) own symbol for it — and heat drains out through two routes — a fixed still-air route (U0: radiation plus natural convection off both faces) and a wind route whose capacity grows with every metre per second (U1·v). The water level — the cell's temperature excess over ambient — rises until the drain matches the tap: exactly S/(U0 + U1v), which is the whole Faiman equation (Fig. 8.1).

The consequence is the curve in Fig. 8.2: under a strong 800 W/m² sun, a still-standard 1 m/s panel runs ≈25 °C above ambient, but at a breezy 5 m/s only ≈13.5 °C — 11.6 °C cooler, worth about 4.6% of output at −0.4%/°C (§9). The curve is steepest at low wind: the first few m/s of breeze do most of the cooling, which is why ignoring wind (as NOCT does) systematically misjudges windy sites.

wind v — heat loss U₁·vcell warms untilinflow = outflowS — sunlight inU₀ — radiation +still-air convection(both faces shed heat)
Fig. 8.1 — The Faiman heat budget. Absorbed sunlight S pours heat in; two routes carry it out — a fixed still-air route (U0: radiation plus natural convection, both faces) and a wind route whose capacity grows with every m/s (U1·v). The cell warms until outflow balances inflow, which is exactly Tcell = Tamb + S/(U0 + U1v): stronger wind, wider drain, cooler panel. Original figure; model per Faiman (2008), IEC 61853-2.
+10+20+300246810wind speed v (m/s)cell temperature above ambient (°C) — at 800 W/m²NOCT assumption: 1 m/s → +25.1 °Cbreezy site: 5 m/s → +13.5 °C
Fig. 8.2 — How much hotter than ambient the cell runs, against wind speed, at a strong-sun 800 W/m² — computed from the model's Faiman constants. The NOCT assumption is the single dot at 1 m/s (+25 °C); a breezy 5 m/s site actually runs 11.6 °C cooler, worth ≈4.6% of output at −0.4%/°C. The curve is steepest at low wind — the first few m/s do most of the cooling.
(8.1) Faiman (2008) — wind-aware cell temperature
Tcell=Tamb+SU0+U1 vw [∘C]{\htmlData{sym=T_cell}{T_{\text{cell}}}} = {\htmlData{sym=T_amb}{T_{\text{amb}}}} + \dfrac{{\htmlData{sym=Sirr}{S}}}{{\htmlData{sym=U_0}{U_0}} + {\htmlData{sym=U_1}{U_1}}\,{\htmlData{sym=v_w}{v_w}}}\ [^\circ\text{C}]
(8.2) Datasheet NOCT → U₀ (per-module thermal tier)
U0=800NOCT−20−U1 [W/(m2K)]{\htmlData{sym=U_0}{U_0}} = \dfrac{800}{{\htmlData{sym=NOCTv}{\text{NOCT}}} - 20} - {\htmlData{sym=U_1}{U_1}}\ [\text{W/(m}^2\text{K)}]
SymbolMeaningUnitsValue / source
TcellModule cell temperature°Coutput
TambAmbient air temperature (lapse-corrected, §3)°CPVGIS T2m (or ERA5 fallback)
SPlane-of-array irradiance — the same quantity as Itotal (§5/§6), in Faiman's notationW/m²per configuration
vwWind speed at 10 m, monthly mean broadcast to all hours of the monthm/sNASA POWER WS10M 2001–2020; clipped to ≥ 0.5 m/s; defaults to 1.0 if unavailable
U0Radiative + natural-convection heat loss at zero windW/(m²·K)default 25.0 — Faiman (2008), c-Si open rack; per-module override, or derived from the datasheet NOCT
U1Wind-driven convective heat lossW·s/(m³·K)default 6.84 — i.e. W/(m²·K) per m/s of wind; per-module override
NOCTDatasheet nominal operating cell temperature°Cmodule parameter (guard 33–58 °C) → U₀ = 800/(NOCT − 20) − U₁; a direct U₀ entry wins

Consistency check: at NOCT reference conditions (800 W/m², 1 m/s, 20 °C) Faiman gives 20 + 800/31.84 ≈ 45.1 °C — matching the standard NOCT of 45 °C. The models agree at the reference point and diverge at every other wind speed. The dashboard's wind thermal gain card is exactly this divergence: annual yield with Faiman at real wind minus annual yield at the NOCT 1 m/s assumption (both computed with your module's thermal coefficients, so the card isolates wind, not module choice).

Your module's thermal behaviour: datasheets state NOCT rather than Faiman coefficients, so the module form runs the same relationship in reverse — NOCT pins the total loss coefficient at the 1 m/s test point, U₀ = 800/(NOCT − 20) − U₁, keeping the wind-dependent share U₁ at the c-Si value unless you enter a measured pair (eq 8.2). A cooler-running module (lower NOCT) gets a larger U₀ and loses less to heat everywhere.

Wind data: NASA POWER MERRA-2 WS10M 20-year climatology (2001–2020), fetched in m/s. WS10M is the wind speed at the standard 10-metre reference height above ground; the model applies no height correction — the 10 m monthly means feed the Faiman equation (8.1) directly, clipped to ≥ 0.5 m/s. A module typically sits below 10 m, where wind is somewhat slower, so the modelled wind cooling is, if anything, slightly optimistic. The 20-year window matters: it spans roughly four full El Niño–Southern Oscillation (ENSO) cycles, so no single phase — such as the 2020–2022 triple La Niña — can bias the wind climatology the way a short window would (§15).

9. PV Power Output & Temperature Derating▾

Concept — hot semiconductors work worse — reversibly

A solar cell is a semiconductor junction, and heat is its enemy twice over — in two different ways this model keeps strictly separate. The instantaneous effect is here in §9: raising the cell temperature lowers the junction voltage, so output falls by |γ| ≈ 0.4% for every °C above the 25 °C rating point — a cell at 65 °C delivers 84% of its rated conversion. This penalty is fully reversible: cool the panel and the power comes straight back, which is why wind (§8) and active cooling (§13) pay off immediately. The cumulative effect — heat permanently ageing the module chemistry — is a separate mechanism with separate physics, modelled in §12; it does not reverse.

(9.1) Power with temperature derating
P=η ⋅ S ⋅ (1+γ (Tcell−25)) ⋅ ηrel(S) [W/m2]{\htmlData{sym=Ppow}{P}} = {\htmlData{sym=eta}{\eta}}\,\cdot\,{\htmlData{sym=Sirr}{S}}\,\cdot\,\big(1 + {\htmlData{sym=gamma}{\gamma}}\,({\htmlData{sym=T_cell}{T_{\text{cell}}}} - 25)\big)\,\cdot\,{\htmlData{sym=eta_rel}{\eta_{\text{rel}}}}({\htmlData{sym=Sirr}{S}})\ [\text{W/m}^2]
(9.2) Low-light relative efficiency (optional datasheet tier)
measured points (IEC 61853-1), Si≤S≤Si+1 — 1 at/above STC:ηrel(S)=ηrel,i+(ηrel,i+1−ηrel,i) ln⁡(S/Si)ln⁡(Si+1/Si)only the 200 W/m2 point printed:ηrel(S)=1+ηrel,200−1ln⁡0.2 ln⁡ ⁣S1000\begin{aligned}&\text{measured points (IEC 61853-1), } {\htmlData{sym=S_i}{S_i}} \le {\htmlData{sym=Sirr}{S}} \le {\htmlData{sym=S_i1}{S_{i+1}}}\text{ — 1 at/above STC:} \\[2pt] {\htmlData{sym=eta_rel}{\eta_{\text{rel}}}}({\htmlData{sym=Sirr}{S}}) &= {\htmlData{sym=eta_rel_i}{\eta_{\text{rel},i}}} + \bigl({\htmlData{sym=eta_rel_i1}{\eta_{\text{rel},i+1}}} - {\htmlData{sym=eta_rel_i}{\eta_{\text{rel},i}}}\bigr)\,\dfrac{\ln({\htmlData{sym=Sirr}{S}}/{\htmlData{sym=S_i}{S_i}})}{\ln({\htmlData{sym=S_i1}{S_{i+1}}}/{\htmlData{sym=S_i}{S_i}})} \\[10pt] &\text{only the 200 W/m}^2\text{ point printed:} \\[2pt] {\htmlData{sym=eta_rel}{\eta_{\text{rel}}}}({\htmlData{sym=Sirr}{S}}) &= 1 + \dfrac{{\htmlData{sym=eta_rel200}{\eta_{\text{rel,200}}}} - 1}{\ln 0.2}\,\ln\!\frac{{\htmlData{sym=Sirr}{S}}}{1000}\end{aligned}
SymbolMeaningUnitsValue / source
PDC power output per unit area (floored at 0)W/m²output; thesis eq 25 expressed per m² (eq 26)
ηModule efficiency at standard test conditions (STC: 1000 W/m², 25 °C)—default 0.20 (typical premium c-Si); with module parameters set, derived from the datasheet: η = rated power ÷ (area × 1000 W/m²), guarded 5–30%
γMaximum-power temperature coefficient/°Cdefault −0.004 (−0.4%/°C, typical crystalline silicon); the datasheet γ_Pmax when module parameters are set
SPlane-of-array irradiance (= Itotal, §5/§6)W/m²from §5/§6
TcellCell temperature°Cfrom §8
ηrelLow-light relative efficiency—the eq 9.2 interpolation through the datasheet 200–800 W/m² anchors when supplied; with none supplied, the technology-typical curve for the chosen cell type (never flat)

Every 1 °C above 25 °C costs ~0.4% of rated output — a desert panel at 65 °C loses 16% of nameplate power purely from heat. This is the instantaneous, reversible thermal effect; the irreversible long-term decline is modelled separately in §12.

Low-light behaviour (optional): real modules convert weak light slightly less efficiently than STC light (series/shunt resistance effects) — and the deviation is not always monotone: many modules peak slightly above STC efficiency around 400–600 W/m². Datasheets quantify this at the IEC 61853-1 test points — 200, 400, 600 and 800 W/m² at 25 °C, e.g. "96.5% of STC at 200 W/m²". Enter whichever points your datasheet gives: eq (9.2) passes exactly through every supplied point (log-linear between neighbours), flat at and above STC and clamped at zero below the characterised range. Interpolating the measured matrix rather than fitting a curve shape is also how the IEC 61853-3 energy rating consumes these measurements; with only the 200 W/m² value this reduces to a one-parameter log fit through the two anchors. Use the 25 °C row only — other rows would double-count the temperature response already modelled by γ (eq 9.1).

No low-light values entered? The model does not assume a flat response — it falls back to a technology-typical curve for the cell type selected above, at the same four points: mono-Si 96/99/100/100%, poly-Si 95/98/99/100%, HJT 95/99/100/100%, CdTe 98/100/100/100% and CIGS 98/99/100/100% at 200/400/600/800 W/m². These come from measurements on commercial modules: the Sandia National Laboratories IEC 61853-1 matrix dataset (Driesse et al. 2020), the CdTe row from Baumgartner et al. 2009. They are typical, not yours — module spread within a technology exceeds the difference between technologies (IEA PVPS T13-20, 2020) — so the datasheet's own points always win when supplied.

What counts as low-light data: the numbers this model consumes are relative-efficiency statements — the module's measured efficiency at each test irradiance as a percentage of its own STC efficiency, printed as a sentence (“relative efficiency at 200 W/m²: 96.5%”), an IEC 61853-1 matrix row, or an efficiency-versus-irradiance chart. Many sheets instead print families of current–voltage curves at several irradiances — a circuit-design illustration (current scales with light, voltage sags logarithmically), not a measured efficiency specification; the penalty is only implied, by reading each curve's maximum-power point off a graphic. The datasheet reader accepts the stated numbers and the efficiency chart, and deliberately never derives values from current–voltage curve graphics. The physics behind the roll-off is the junction's shunt resistance: in weak light the photocurrent is small while the internal leakage (~V/Rsh) barely shrinks, so a larger share is lost internally — the mirror of series resistance, whose I²R losses bite at high light (De Soto, Klein & Beckman 2006). That is also why the §9.3 curve model refines Rsh from exactly these measured points.

Module electrical model — the full current–voltage curve (optional): eq 9.1 treats the module as one efficiency number: power linear in light, linear in temperature. Enter the datasheet's electrical block — open-circuit voltage Voc, short-circuit current Isc, the rated maximum power point (Vmp, Imp) and the number of cells in series — and the model additionally builds the single-diode model (De Soto, Klein & Beckman 2006, Solar Energy 80:78–88), the standard cell physics the bankable design tools (NREL's System Advisor Model and its peers) are built on. The module's whole current–voltage curve is reconstructed and solved at its maximum power point (MPP) for every hour of the year, on both configurations — so voltage and current become outputs in their own right: the voltage-window statistics and string-sizing check in the results. The energy itself stays on eq 9.1 with your datasheet's measured temperature coefficient — the curve adds the electrical outputs, it does not replace the power calculation.

Concept — a cell is a current source — with a diode stealing the current back

Light makes current: every absorbed photon frees a charge carrier, so the photocurrent IL scales with irradiance — nearly linearly, which is all eq 9.1 assumes. But the cell is also a p–n junction, and a junction is a diode: as the terminal voltage rises, the diode internally consumes more and more of that current, until at the open-circuit voltage it consumes everything and the terminals deliver nothing. Between the two extremes sits the knee of the curve — the maximum power point. Two resistances complete the physics: the series resistance Rs (contacts, silicon bulk) drops voltage under load and rounds the knee, while the shunt resistance Rsh (microscopic leakage paths) bleeds current — the reason efficiency sags at dawn and dusk.

Five numbers from the datasheet pin the curve's five physical parameters (IL, the diode saturation current I0, Rs, Rsh and the modified ideality a): the two endpoints (Isc, Voc), the rated point (Vmp, Imp), and the voltage temperature coefficient βVoc. Solving those five constraints is the extraction. The parameters are then translated to each hour's irradiance and cell temperature — current scales with light, while I0 climbs steeply with temperature (through the semiconductor bandgap, per cell technology) — and it is that climb which drags the voltage, and with it the power, down on hot days. The linear γ of eq 9.1 is the straight-line approximation to this mechanism; the curve model computes the mechanism itself.

absorbed light SIL — photocurrent(light makes current)junction diode(steals current back as V rises)Rsh — shunt(leakage paths)Rs — series+−VIto the load
Fig. 9.1 — The single-diode equivalent circuit. Absorbed light acts as a current source IL; the cell's own p–n junction is the diode in parallel, internally consuming more and more of that current as the terminal voltage rises (all of it at Voc, none in the dark); the shunt resistance Rsh leaks a further share, and the series resistance Rs drops voltage between the junction and the terminals. What the load actually receives is what the diode and shunt do not take — the single-diode equation is exactly that split, and its five parameters come from the five datasheet constraints. Original figure; topology per De Soto, Klein & Beckman (2006).

Concept — the fit can only be as good as the datasheet block

Datasheet values are rounded to two or three significant figures, and five rounded numbers cannot always be hit exactly by any single-diode curve. The model therefore fits in tiers. First an exact fit: all five constraints satisfied to a fraction of a per cent. When no exact curve exists, it falls back to the relaxed fit the System Advisor Model uses (the CEC structure; Dobos 2012): short-circuit current and open-circuit voltage are matched exactly, and the rated voltage/current point is reproduced approximately — with the fit mode and the size of that difference stated openly in the results, never silently absorbed. The relaxed fit is the industry-standard resolution of rounded datasheet data, not a shortcut.

Two further relaxations are reported the same way. Datasheets often state a temperature coefficient steeper or shallower than any single-diode curve with that open-circuit voltage can physically produce — the fit then clamps it to the physics boundary and states the adjustment, the same correction the System Advisor Model applies silently (its "Adjust" column). And if the nameplate power exceeds what the curve can deliver at any series resistance, the fit takes the curve's own maximum and reports the shortfall. A large reported adjustment is your warning light: the usual cause is the half-cell trap — a modern "144-cell" module is 144 half-cells, i.e. 72 cells wired in series, and entering 144 mis-states every per-cell voltage. The simulation refuses only when the numbers are structurally impossible — a rated maximum-power point at or beyond the open-circuit voltage or short-circuit current — which is mis-transcription, not physics. (Per-technology medians pre-fill so an unknown coefficient never blocks you, but a supplied datasheet value always wins.)

(9.3) Single-diode model (De Soto 2006)
I=IL−I0[exp⁡ ⁣(V+I Rsa)−1]−V+I RsRsha=n Ns kB Tcellq [V]{\htmlData{sym=I_mod}{I}} = {\htmlData{sym=I_L}{I_L}} - {\htmlData{sym=I_0}{I_0}}\left[\exp\!\left(\dfrac{{\htmlData{sym=V_mod}{V}} + {\htmlData{sym=I_mod}{I}}\,{\htmlData{sym=R_s}{R_s}}}{{\htmlData{sym=a_mod}{a}}}\right) - 1\right] - \dfrac{{\htmlData{sym=V_mod}{V}} + {\htmlData{sym=I_mod}{I}}\,{\htmlData{sym=R_s}{R_s}}}{{\htmlData{sym=R_sh}{R_{sh}}}} \qquad {\htmlData{sym=a_mod}{a}} = \dfrac{{\htmlData{sym=n_diode}{n}}\,{\htmlData{sym=N_s}{N_s}}\,{\htmlData{sym=kb}{k_B}}\,{\htmlData{sym=T_cell}{T_{\text{cell}}}}}{{\htmlData{sym=q_e}{q}}}\ \text{[V]}
(9.4) Extraction — the five datasheet constraints, fitted in tiers
I(0)=IscI(Voc)=0I(Vmp)=ImpdPmpdV∣Vmp=0dVocdTcell=βVoc\begin{aligned} {\htmlData{sym=I_mod}{I}}(0) &= {\htmlData{sym=I_sc}{I_{\text{sc}}}} \qquad {\htmlData{sym=I_mod}{I}}({\htmlData{sym=V_oc}{V_{\text{oc}}}}) = 0 \qquad {\htmlData{sym=I_mod}{I}}({\htmlData{sym=V_mp}{V_{\text{mp}}}}) = {\htmlData{sym=I_mp}{I_{\text{mp}}}} \\ \left.\dfrac{{\htmlData{sym=d_diff}{\mathrm{d}}}{\htmlData{sym=P_mp}{P_{\text{mp}}}}}{{\htmlData{sym=d_diff}{\mathrm{d}}}{\htmlData{sym=V_mod}{V}}}\right|_{{\htmlData{sym=V_mp}{V_{\text{mp}}}}} &= 0 \qquad \dfrac{{\htmlData{sym=d_diff}{\mathrm{d}}}{\htmlData{sym=V_oc}{V_{\text{oc}}}}}{{\htmlData{sym=d_diff}{\mathrm{d}}}{\htmlData{sym=T_cell}{T_{\text{cell}}}}} = {\htmlData{sym=beta_Voc}{\beta_{V_{\text{oc}}}}} \end{aligned}
(9.5) Translation to each hour's irradiance and temperature
IL=SeSref(IL,ref+αIsc(Tcell−Tref))I0=I0,ref(TcellTref) ⁣3exp⁡ ⁣[1n(Eg,refkB Tref−Eg(Tcell)kB Tcell)]\begin{aligned} {\htmlData{sym=I_L}{I_L}} &= \dfrac{{\htmlData{sym=S_e}{S_e}}}{{\htmlData{sym=S_ref}{S_{\text{ref}}}}}\bigl({\htmlData{sym=I_Lref}{I_{L,\text{ref}}}} + {\htmlData{sym=alpha_Isc}{\alpha_{I_{\text{sc}}}}}({\htmlData{sym=T_cell}{T_{\text{cell}}}} - {\htmlData{sym=T_ref}{T_{\text{ref}}}})\bigr) \\ {\htmlData{sym=I_0}{I_0}} &= {\htmlData{sym=I_0ref}{I_{0,\text{ref}}}}\left(\dfrac{{\htmlData{sym=T_cell}{T_{\text{cell}}}}}{{\htmlData{sym=T_ref}{T_{\text{ref}}}}}\right)^{\!3}\exp\!\left[\dfrac{1}{{\htmlData{sym=n_diode}{n}}}\left(\dfrac{{\htmlData{sym=E_gref}{E_{g,\text{ref}}}}}{{\htmlData{sym=kb}{k_B}}\,{\htmlData{sym=T_ref}{T_{\text{ref}}}}} - \dfrac{{\htmlData{sym=E_g}{E_g}}({\htmlData{sym=T_cell}{T_{\text{cell}}}})}{{\htmlData{sym=kb}{k_B}}\,{\htmlData{sym=T_cell}{T_{\text{cell}}}}}\right)\right] \end{aligned}
(9.6) Hourly operating point — maximum power
Pmp=max⁡V [ V⋅I(V) ]solved wheredPmpdV=I+V dIdV=0{\htmlData{sym=P_mp}{P_{\text{mp}}}} = \max_{{\htmlData{sym=V_mod}{V}}}\ \bigl[\, {\htmlData{sym=V_mod}{V}} \cdot {\htmlData{sym=I_mod}{I}}({\htmlData{sym=V_mod}{V}}) \,\bigr] \qquad \text{solved where} \qquad \dfrac{{\htmlData{sym=d_diff}{\mathrm{d}}}{\htmlData{sym=P_mp}{P_{\text{mp}}}}}{{\htmlData{sym=d_diff}{\mathrm{d}}}{\htmlData{sym=V_mod}{V}}} = {\htmlData{sym=I_mod}{I}} + {\htmlData{sym=V_mod}{V}}\,\dfrac{{\htmlData{sym=d_diff}{\mathrm{d}}}{\htmlData{sym=I_mod}{I}}}{{\htmlData{sym=d_diff}{\mathrm{d}}}{\htmlData{sym=V_mod}{V}}} = 0
(9.7) Substrings and the bypass-diode clamp
per substring:Ns/Nsub cells,Rs/Nsub,  Rsh/Nsub,  a/Nsubshaded substring clamps at −Vf,  Vf=0.6 V\begin{aligned} &\text{per substring:} \quad {\htmlData{sym=N_s}{N_s}}/{\htmlData{sym=N_sub}{N_{\text{sub}}}}\ \text{cells}, \quad {\htmlData{sym=R_s}{R_s}}/{\htmlData{sym=N_sub}{N_{\text{sub}}}},\ \ {\htmlData{sym=R_sh}{R_{sh}}}/{\htmlData{sym=N_sub}{N_{\text{sub}}}},\ \ {\htmlData{sym=a_mod}{a}}/{\htmlData{sym=N_sub}{N_{\text{sub}}}} \\ &\text{shaded substring clamps at} \ {-}{\htmlData{sym=V_f}{V_f}},\ \ {\htmlData{sym=V_f}{V_f}} = 0.6\ \text{V} \end{aligned}
SymbolMeaningUnitsValue / source
ILLight-generated photocurrent — scales with irradianceAextracted from the datasheet block; translated hourly
I0Diode saturation current — climbs steeply with temperature via the bandgapAextracted; translated hourly
RsSeries resistance — contact and bulk losses; rounds the MPP kneeΩextracted; constant with operating conditions
RshShunt resistance — leakage paths; shapes the low-light roll-offΩextracted; refined toward the datasheet low-light points when supplied
aModified ideality n·Ns·kB·T/q — diode knee sharpnessVextracted; scales with temperature
nDiode ideality factor (junction quality)—extracted; the relaxed fit searches the physical 0.8–2.5 range
NsCells in series—datasheet — beware half-cut cells: 144 half-cells = 72 in series
EgSemiconductor bandgapeVper technology: c-Si 1.121 (De Soto 2006), CdTe 1.5, CIGS 1.15
αIscShort-circuit-current temperature coefficient%/°Cdatasheet wins; per-technology median pre-fills (c-Si +0.05)
βVocOpen-circuit-voltage temperature coefficient%/°Cdatasheet wins; median pre-fills; clamped to the physics envelope, adjustment reported
NsubSubstrings — equal cell groups, one per bypass diode—datasheet diode count; must divide the series count evenly
VfBypass-diode forward voltage — clamps a shaded substringV0.6 (Schottky); dormant under uniform light

What the curve provides — and what it deliberately does not: the yield always runs eq 9.1 on the γ coefficient and low-light anchors your datasheet states — those are the manufacturer's measured values, and they stay in charge of the energy. The curve provides what only a curve can: the voltage windows, the STC figure, and a consistency diagnostic — an implied γ (from curve power at 25 °C versus 50 °C) and implied relative efficiency at the IEC low-light points, each reported against your datasheet values. A sizeable gap between the implied and datasheet γ is expected for modern low-γ modules: the single-diode temperature physics is anchored to a fixed textbook silicon bandgap that cannot reproduce today's gentler printed coefficients — which is exactly why the datasheet value, not the curve's, drives the energy.

Substrings and bypass diodes: a real module wires its cells in equal substrings, each bridged by a bypass diode that conducts around its group if that group can no longer carry the string current — the partial-shading safety mechanism. Under uniform light the diodes never conduct (the model reproduces the whole-module curve exactly either way); the architecture is in place so the shading model, when it arrives, can activate them (see What Is Not Yet Modelled).

Bifacial modules: rear-side light enters as extra effective irradiance (§17), which the curve converts into extra current — the bifacial gain is physically a current gain, so the electrical layer needs no special handling and inherits it exactly.

Fallback: without the full electrical block (all four standard-test-condition values plus the series cell count), everything stays on the straight-line model of eq 9.1 exactly as before — the two paths are pinned identical by test when the block is absent.

10. Dust & Soiling Model▾

Soiling rate (Kimber 2006 adapted; Micheli & Müller 2017 calibration): dust deposits at a daily rate proportional to the atmospheric aerosol load:

Concept — a dust gauge in orbit, and a sawtooth on the ground

No one measures dust on your panel, so the model uses the next best thing: the aerosol optical depth (AOD) — how much a satellite's view of the ground is dimmed by everything suspended in the air column. Hazier air both blocks sunlight and settles onto glass, so AOD is a physically motivated proxy for how fast panels get dirty: the deposition rate is taken as proportional to it (eq 10.1), with a tilt correction (flat panels catch more) and the user multiplier for local dust a ~55 km satellite pixel cannot see.

Between cleanings, the loss traces a sawtooth (Fig. 10.1): dust builds linearly, then a cleaning visit — or ≥5 mm of rain — resets it to zero. The model never simulates day-by-day; it uses the closed-form mean of one tooth: peak/2 for a triangle, switching to the flat-top trapezoid form once the build-up hits the Smax = 30% saturation ceiling, where fresh dust stops adding loss (it stacks and sheds rather than covering new glass).

10%20%30%dashed: mean = peak/2 (5%)clean / rainlight soiling — triangular (1%/day, 10-day cycle)saturation cap — losses stop growingtrapezoid mean = 21%day 15: cap reachedheavy soiling — flat-top trapezoid (2%/day, 25-day cycle)
Fig. 10.1 — Dust builds linearly at the soiling rate SR and resets to zero at each cleaning — scheduled or rain. Left: while the peak stays under the cap, the sawtooth is triangular and the month's mean loss is simply peak/2 (here SR = 1%/day cleaned every 10 days → mean 5%). Right: heavy soiling saturates — beyond Smax = 30% new dust stops adding loss (it sheds and stacks rather than covering new glass; Ferrada et al. 2018) — so the tooth becomes a flat-top trapezoid and the mean follows eq (10.2): here SR = 2%/day over 25 days caps at day 15, mean 21%. Original figure.

Concept — dust hurts twice — blockage plus reflection

A dust film costs energy through two distinct physical channels (Fig. 10.2). The obvious one is blockage: grains absorb and shade — that is the s̄ term, capped at 30%. The subtler one is reflection: a dusty surface scatters more of an oblique beam away before it ever enters the glass, which in the Martín–Ruiz language means αr rises with soiling (0.17 clean → 0.27 at the cap) and the whole IAM curve of Fig. 6.3 sags.

The two compound (Fig. 10.3): head-on, a panel at the 30% cap still delivers 70% of its clean beam — but at θ = 65° it delivers only ≈62%, because the same dust now also reflects. The extra penalty grows fastest between 20% and 30% soiling, exactly where αr climbs most steeply. Note who pays it: a fixed panel spends every morning and evening at large θ, while a tracker holds θ ≈ 0 — so soiling widens the tracking gain, and the combined formula in 9.3 below is applied to the fixed-tilt side only (two-axis IAM stays ≈1 at any αr). It is also built to avoid double counting: the yield is interpolated at αr,eff first, then scaled by the blockage factor.

dust filmglass123① reflected — αᵣ rises with dust · ② blocked by the film — s̄ (≤ 30%) · ③ transmitted — (1 − s̄) × IAM
Fig. 10.2 — Three identical rays meet a dusty panel, and each shows one fate. ① Reflected: a dust layer scatters extra light away at every angle — this is what raises the Martín–Ruiz αr from 0.17 toward 0.27 and drags the whole IAM curve down (Fig 4.3). ② Blocked: the film directly shades the cell — the s̄ term, which saturates at Smax = 30% (Fig 9.1). ③ Transmitted: what the cell receives — (1 − s̄) × IAM(αr,eff), the two channels compounding (Fig 9.3). The model combines them without double counting: loss = yield − yield(αr,eff) × (1 − s̄). Original figure.
60%70%80%90%100%0%5%10%15%20%25%30%mean soiling level s̄beam delivered, relative to a clean panel at the same angleθ = 0° — blockage only (1 − s̄)θ = 65° — blockage + extra reflectionextra angular penalty
Fig. 10.3 — The same dust costs more at oblique angles — computed from the model's own αr(s̄) map and Martín–Ruiz formula, each curve relative to a clean panel at the same angle. Head-on (θ = 0°), dust costs exactly its blockage: 30% soiling → 70% of clean output. At θ = 65° the raised αr adds a second penalty on top, leaving ≈62% — the shaded wedge is pure angular loss, and it widens fastest between 20% and 30% soiling, where αr climbs most steeply (0.20 → 0.27). Fixed-tilt panels spend their mornings and evenings at exactly these angles; a tracker at θ ≈ 0 never pays the second penalty. Original figure.
(10.1) Soiling rate (Kimber 2006 · Micheli 2017)
SR=Ksoil ⋅ AOD550 ⋅ cos⁡Σ ⋅ fuser [%/day]{\htmlData{sym=SR}{\text{SR}}} = {\htmlData{sym=K_soil}{K_{\text{soil}}}}\,\cdot\,{\htmlData{sym=AOD}{\text{AOD}_{550}}}\,\cdot\,\cos{\htmlData{sym=Sigma}{\Sigma}}\,\cdot\,{\htmlData{sym=f_user}{f_{\text{user}}}}\ [\%/\text{day}]
(10.2) Mean soiling over a cleaning cycle
s‾={p/2p≤SmaxSmax (1−ncap/(2d))p>Smax  [%]{\htmlData{sym=sbar}{\overline{s}}} = \begin{cases} {\htmlData{sym=p_peak}{p}}/2 & {\htmlData{sym=p_peak}{p}} \le {\htmlData{sym=S_max_v}{S_{\text{max}}}} \\[2pt] {\htmlData{sym=S_max_v}{S_{\text{max}}}}\,\big(1 - {\htmlData{sym=n_cap_v}{n_{\text{cap}}}}/(2{\htmlData{sym=d_int}{d}})\big) & {\htmlData{sym=p_peak}{p}} > {\htmlData{sym=S_max_v}{S_{\text{max}}}} \end{cases}\ \ [\%]
(10.3) Combined loss (optical + IAM)
net=yield(αr,eff) ⋅ (1−s‾/100) [kWh/m2]{\htmlData{sym=netf}{\text{net}}} = {\htmlData{sym=yieldf}{\text{yield}}}({\htmlData{sym=alpha_reff}{\alpha_{r,\text{eff}}}})\,\cdot\,(1 - {\htmlData{sym=sbar}{\overline{s}}}/100)\ [\text{kWh/m}^2]
SymbolMeaningUnitsValue / source
SRDaily soiling rate (loss accumulation speed)%/dayoutput of (9.1)
KsoilSoiling rate coefficient%·day⁻¹·AOD⁻¹0.65 — Micheli & Müller (2017) / pvlib Kimber calibration
AOD550Aerosol optical depth at 550 nm — satellite dust-load proxy—NASA POWER SYN1DEG AOD_55, monthly means 2018–2022; missing months fall back to the annual mean (flagged)
cos ΣTilt correction — steeper panels accumulate less dust—Coello & Boyle (2019), IEEE J. Photovolt. 9:1382; accumulation only — rain cleaning is tilt-agnostic (known simplification)
fuserUser dust multiplier for local sources invisible to satellite—×1 / ×3 / ×8 / ×15
SmaxSaturation ceiling — losses stop growing%30 — Ferrada et al. (2018) Atacama field data
RMonthly rainfall totalmm/monthNASA POWER PRECTOTCORR 2001–2020 (converted from mm/day)
nrainEffective natural cleaning events per month—R/10 when R ≥ 5 mm, else 0 (Kimber threshold; 10 mm ≈ one cleaning event)
dEffective cleaning interval for a scenariodaystable below
pPeak soiling at the end of a cycle%SR·d
s̄Mean soiling level over an accumulate-then-reset cycle%triangular / trapezoidal average, (9.2)
yieldFixed-tilt monthly yield as a function of αr, interpolated at αr,effkWh/m²from the precomputed per-αr tables (§6/§7)
netFixed-tilt yield after the combined soiling losskWh/m²yield(αr,eff) × (1 − s̄/100), (9.3)

Accumulation & averaging (statistics): soiling builds linearly at SR %/day and resets to zero at each cleaning event, producing a sawtooth. The mean loss over a cycle of d days has a closed form: peak/2 (triangular average) while the peak SR·d stays below Smax; once it saturates, the sawtooth becomes a flat-top trapezoid with mean Smax·(1 − ncap/2d), where ncap = Smax/SR is the number of days to reach the cap.

Effective cleaning intervals (n = days in the month): rainfall enters the model here — each 10 mm of monthly rainfall R counts as one natural cleaning event, nrain = R/10 for R ≥ 5 mm (below the Kimber threshold rain contributes no cleaning). Scheduled cleans and rain events between visits combine, so a manual schedule never does worse than doing nothing. In a dry month the intervals collapse to exactly 1 / 7 / n days.

ScenarioInterval d (days)Reading
Perfect0continuously clean — zero loss
Dailyn / (n + nrain)≈ 1 day
Weeklyn / (n/7 + nrain)weekly visits + rain in between
Monthlyn / (1 + nrain)monthly visit + rain in between
No actionmin(n / max(nrain, 0.5), n)rain-only; floored at 0.5 events/month (wind and dew provide a baseline reset) and capped at the month length

IAM angular correction (9.3): dust also elevates the Martín–Ruiz αr parameter — dusty glass reflects more at oblique angles, separately from the light it blocks. αr maps piecewise-linearly from mean soiling (0% → the module's clean αr (0.17 generic, 0.14 AR-coated glass), then 10% → 0.19, 20% → 0.20, 30% → 0.27 — the soiled points are dust-dominated and module-independent; Martín & Ruiz 2005) and monthly yield is interpolated from the precomputed per-αr tables (§6/§7), whose clean point shifts with the module too — an AR module shows no spurious loss at 0% soiling. The combined loss — angular penalty and direct blockage without double-counting — is loss = yield − yield(αr,eff) × (1 − s̄/100). The two-axis beam is unaffected (at θ ≈ 0 the beam IAM is 1 for any αr); its diffuse/ground factors do share αr (§5), but the soiling loss itself is evaluated on the fixed-tilt reference. For a bifacial module (§17) both sides of this difference are front-only monofacial yields — dust settles on the upward glass, so the rear bonus can neither soil nor be booked as a soiling loss — while the reported percentages divide by the configuration's full bifacial yield.

Mine-site limitations: Ksoil = 0.65 is calibrated against background atmospheric aerosols at non-industrial sites. A thorough literature review (2026-06) found zero published papers that re-derive this coefficient for operating mine environments. Two errors partially cancel in an unknown direction: (1) MERRA-2 AOD at ~55 km grid resolution cannot see a single mine pit — locally generated dust from vehicle haulage, blasting, and processing is invisible to satellite data; (2) coarse mine particles have lower optical potency per gram than the fine aerosols that calibrate K globally (Ilse et al. 2021). The user multiplier compensates for the missing local source but remains an engineering estimate, not a calibrated physical coefficient — no published validation for any mine environment exists (IEA PVPS Task 13, 2022/2025). Cementation of dust by wet-dry cycling (gypsum and halite acting as binding agents, confirmed within 2–7 days in arid mine environments — Atacama studies 2019, 2025) is not modelled; cleaning events in the model assume full yield recovery, which may overstate recovery for heavy industrial dust.

11. Snow Model▾

Marion et al. (2013) NREL physical threshold model, adapted to monthly resolution. A snow-covered panel clears when irradiance provides enough thermal energy to overcome bonding forces; the monthly adaptation captures this through a temperature factor, a tilt-driven slide rate, and a daylight-availability scale.

Concept — why 0 °C is worse than −20 °C

Snow leaves a panel mainly by sliding: sunlight warms the dark glass, a melt film forms underneath, and the slab skis off under its own weight. Tilt sets the pace — the Marion slide rate says a full clear takes 1/(0.197 sin Σ) hours of sliding conditions: ≈29 h at a 10° tilt, 9.6 h at 32°, 5.9 h at 60°. Steep panels shed; near-flat panels hold their snow for days.

Temperature decides whether sliding is possible, and the answer is not monotonic (Fig. 11.1). Deep cold means dry, low-friction powder that slides readily; warmth means outright melting. The worst case sits in between: near 0 °C, daytime melt refreezes into an ice bond welding the slab to the glass, and clearing takes nearly twice the nominal time. A panel at −20 °C and one at +5 °C both clear faster than one at 0 °C — which is why mild, marginal snow climates can lose more yield per centimetre of snowfall than genuinely cold ones.

0.5×1.0×1.5×2.0×-20-15-10-30210monthly mean temperature (°C)clearing-time multiplier (higher = snow stays longer)dry powder — slides freelymelt–refreeze ice bond — slowestmelts off
Fig. 11.1 — The temperature factor fT multiplies the snow-clearing time, straight from the model's code. The peak is the counterintuitive part: at −20 °C snow is dry low-friction powder that slides readily (0.6×), and above +10 °C it simply melts (0.4×) — but near 0 °C partial melt refreezes into an ice bond with the glass, and clearing takes nearly twice the nominal time (1.8×). A panel at −20 °C and one at +5 °C both clear faster than one at 0 °C. Marion et al. (2013) bonding physics, adapted to monthly resolution.

Concept — the tracker's escape hatch

Everything above conspires against a shallow, fixed panel: sin Σ is small, so sliding is slow; it cannot move, so it waits out the ice-bond regime at whatever tilt it was built with. A two-axis tracker holds the one lever that matters — it can rotate to Σ ≈ 85–90°, where the slide physics runs at its fastest (≈5 h to a full clear versus ≈20 h at a 15° tilt), while its dark face keeps absorbing sunlight, feeding the melt film the slab slides on (Fig. 11.2). For this reason the model applies snow losses to the fixed-tilt configurations only — the two-axis yield carries none, an idealisation of a tracker with a steep-stow snow strategy. (The perfect scenario plays the same role for fixed panels: heated or continuously cleared, zero coverage.)

fixed panel — Σ = 15°full clear ≈ 20 h of sliding — days once the ice bond bitesbarely slidesstows near-verticaldark face absorbs sun →melt film lubricates the slidetwo-axis tracker — stowed at Σ ≈ 85–90°full clear ≈ 5 h — the fastest the slide physics allows
Fig. 11.2 — Tilt is the snow-shedding lever, and only the tracker can pull it. The Marion slide rate scales with sin Σ: a shallow 15° fixed panel needs ≈20 h of sliding conditions for one full clear — with the near-freezing bond factor (fT = 1.8, Fig 10.1) that stretches to days, so slabs sit through exactly the season the panel needs most. A two-axis tracker simply rotates to ≈85–90°, where the same physics gives ≈5 h — the fastest clear the model allows — while its dark face soaks up sun, feeding the melt film the slab slides on. This is why the model applies snow losses to the fixed-tilt configurations only. Original figure.

Concept — how hours of physics become days of coverage

The clearing-time formula divides two different kinds of time, and avail is the exchange rate between them. The numerator, slidehours·fT, is in hours of sliding conditions — hours with enough sun on the glass to keep the melt film alive. But the model needs its answer in days of coverage, and a day never offers 24 such hours: the melt film is solar-powered, so the clock stops at night and barely runs under heavy overcast. The denominator, 8·avail, is the conversion — how many sliding-capable hours one day of this particular month supplies. Dividing hours by hours-per-day leaves tclear in days.

The reference window is 8 h/day — roughly the useful-sun hours of a clear winter day — and avail is the fraction of that window the month actually delivers. The model infers it from the month's own fixed-tilt yield, the most direct evidence it has of how much sun reached the panel: 40 kWh/m² over a month is taken as what the full window would produce, so avail = clip(yield / 40, 0.15, 1). A month at avail = 1 supplies the whole 8-hour window; a gloomy month at avail = 0.5 supplies only ≈4 useful hours a day, so the same clear takes twice as many days. avail carries no units because it is a like-for-like ratio — this month's melt-driving sun over a reference month's.

The clips are guard rails rather than physics. The 0.15 floor (≈1.2 h/day) stops dim-but-not-dark months from stretching tclear to absurd lengths, and the cap at 1 stops bright months from clearing faster than the slide physics itself allows — 8 h of sliding conditions is already all a winter day can usefully offer a slab. Below 0.3 kWh/m² for the whole month the model stops estimating altogether: tclear = 999 days, the polar-night guard — no daylight, no melt film, the snow waits for the sun. Both the 8-h window and the 40 kWh/m² reference are calibration choices of this monthly adaptation, not Marion's — the original hourly model never needs them, because it checks the sliding condition against actual irradiance hour by hour (Marion et al. 2013).

(11.1) Snow slide-off time (Marion 2013)
slidehours=10.197 sin⁡Σ [h]{\htmlData{sym=slide_hours}{\text{slide}_{\text{hours}}}} = \dfrac{1}{0.197\,\sin{\htmlData{sym=Sigma}{\Sigma}}}\ [\text{h}]
(11.2) Natural clearing time per event
tclear=slidehours ⋅ fT8 ⋅ avail  [days]{\htmlData{sym=t_clear_v}{t_{\text{clear}}}} = \dfrac{{\htmlData{sym=slide_hours}{\text{slide}_{\text{hours}}}}\,\cdot\,{\htmlData{sym=f_T}{f_T}}}{8\,\cdot\,{\htmlData{sym=availf}{\text{avail}}}}\ \ \text{[days]}
(11.3) Coverage fraction → monthly loss
cover=min⁡ ⁣(nevents⋅min⁡(tclear, I) / days,  1){\htmlData{sym=coverf}{\text{cover}}} = \min\!\big({\htmlData{sym=n_events}{n_{\text{events}}}}\cdot\min({\htmlData{sym=t_clear_v}{t_{\text{clear}}}},\ {\htmlData{sym=I_man}{I}})\,/\,{\htmlData{sym=daysf}{\text{days}}},\ \ 1\big)
(11.4) Ground-snow albedo bonus (attribution)
bonus=yield ⋅ 1−cos⁡Σ2 ⋅ max⁡(ρsnow−ρclear, 0) [kWh/m2]{\htmlData{sym=bonusf}{\text{bonus}}} = {\htmlData{sym=yieldf}{\text{yield}}}\,\cdot\,\dfrac{1-\cos{\htmlData{sym=Sigma}{\Sigma}}}{2}\,\cdot\,\max({\htmlData{sym=rho_snow}{\rho_{\text{snow}}}} - {\htmlData{sym=rho_clear}{\rho_{\text{clear}}}},\ 0)\ [\text{kWh/m}^2]
SymbolMeaningUnitsValue / source
slidehoursHours of continuous sliding to fully clear the panelh1 / (kslide·sin Σ); kslide = 0.197 h⁻¹ (Marion 2013); tilt clamped ≥ 1°
fTTemperature factor — snow-to-glass bonding—0.6 below −15 °C (dry powder) · 1.0 to −3 °C · 1.8 near freezing (−3 to +2 °C, peak ice bond) · 0.8 to +10 °C · 0.4 above (melt-dominated)
availDaylight availability — fraction of the 8 h/day reference sliding window the month actually supplies—clip(monthly yield / 40 kWh/m², 0.15, 1); 40 kWh/m² ≈ what the full window delivers over a month
tclearDays for one snow event to clear naturallydaysslidehours·fT / (8·avail); floored at 0.08 d (~2 h); set to 999 d when monthly yield < 0.3 kWh/m² (polar night — no daylight, no clearing)
neventsSnowfall events per month—monthly snowfall / 5 cm water-equivalent per event
IManual clearing interval per scenariodays∞ (no action) / 30 / 7 / 1 / 0 (perfect)
coverFraction of the month the panel is snow-covered—min(nevents × min(tclear, I) / days, 1); monthly loss = yield × cover
daysDays in the month being evaluateddayscalendar month length
bonusGround-snow albedo bonus yieldkWh/m²(10.4); attribution estimate — the snow-lifted albedo is embedded in the hourly yields; applies in months with > 1 cm water-equivalent snowfall
ρsnowGround albedo with snow cover—0.65
ρclearGround albedo without snow cover—the §2 site albedo (default 0.20) — the bonus is the step up to snow, max(ρ_snow − ρ, 0), so a bright site gains little

Five scenarios — no action / monthly / weekly / daily clearing / perfect (heating cable or continuous mechanical removal — zero coverage) — across all four tilt configurations. Near-freezing months are the worst case: partial melt refreezes into an ice bond, so panels at −20 °C (dry powder) and +5 °C (melt) both clear faster than at 0 °C.

Snow albedo bonus: snow on the ground raises reflected irradiance (site ρ lifted to 0.65). This lift happens inside the hourly simulation — every ground-reflected term, both configurations, front and rear (§17) — in months with more than 1 cm water-equivalent snowfall. The bonus figure reported here is a view-factor estimate of the fixed-tilt front's share of that embedded brightening, shown so the effect stays attributable; it is never added on top. The estimate is computed in broadband albedo — the simulated yields additionally carry the §2 snow-row spectral factor, which this figure deliberately omits (a few per cent of the bonus, inside the estimate's own tolerance). The model flags a site as snow-affected only when annual snowfall exceeds 1 cm water-equivalent.

Snowfall data: Open-Meteo ERA5 snowfall_sum, 2004–2022 (19 years, ~3 ENSO cycles), fetched as daily cm water-equivalent and averaged to mean-daily values per calendar month; the model multiplies by days-in-month for monthly totals. (1 cm water-equivalent ≈ 10 cm of fresh snow depth.)

12. Long-Term Degradation — Arrhenius Model▾

Jordan & Kurtz (2013/2016) NREL model. The default baseline is the crystalline-silicon 0.5%/year at 25 °C — the median of more than 11,000 field measurements. With module parameters set, the baseline resolves per module: your datasheet's warranted degradation rate wins; otherwise a per-technology median (mono 0.5 / poly 0.6 %/yr from the published x-Si range; HJT ≈ 1; CdTe/CIGS 1.0 %/yr as a deliberately conservative fallback — their published medians are confounded, so prefer the warranted figure; Jordan et al. 2016). Heat accelerates the underlying chemistry (encapsulant discolouration, solder fatigue, potential-induced degradation) following an Arrhenius law, so the rate is climate-adjusted from the site's full hourly cell-temperature history.

Year 1 carries a one-off first-year drop in warranty semantics — light-induced degradation (LID) plus any other year-1 effect — that replaces the model's former fixed 2% constant, never stacking on it. Your datasheet's first-year figure wins (0–3%; 0 is a genuine claim for some newer cell technologies); otherwise a per-technology warranty-class ceiling (2% for mono/poly c-Si, CdTe and CIGS; 1% for HJT); otherwise the generic 2% default. The ceilings are warranty posture, not measured medians — measured first-year losses run 0.4–0.7% across cell types (Kiwa PVEL 2026 Module Scorecard) — so the figure printed on your own sheet is always the better input.

Concept — heat is a chemical accelerant — and averages lie

The slow decline of a module is chemistry: encapsulant browning, solder fatigue, ion migration. Like nearly all chemistry, it follows the Arrhenius law — the rate roughly doubles for every 10 °C (with the model's Ea = 0.6 eV: 2.1× from 25 → 35 °C). Rust never sleeps, but heat makes it sprint.

The exponential shape has a consequence that justifies the whole Teq machinery (Fig. 12.1): the curve is convex, so damage evaluated at the average temperature understates the average damage. A module spending half its hours at 25 °C and half at 65 °C sits at 45 °C on average — but its true damage rate is what a constant 54.9 °C would do, nearly double the steady-45 °C rate. Cool nights cannot repay hot afternoons. Teq is exactly this correction, computed over all 8,760 hourly cell temperatures — and it is why wind (§8), altitude (§3) and active cooling (§13) all show up again in the 25-year numbers, not just the instant ones.

5×10×15×025456575cell temperature (°C)damage acceleration factor AF (relative to 25 °C)half 25° / half 65° hours: AF ≈ 8.4steady 45 °C: AF ≈ 4.3
Fig. 12.1 — The Arrhenius acceleration factor, from the model's own constants (Ea = 0.6 eV): damage chemistry roughly doubles every 10 °C, so the curve is convex — and convexity means averages lie. A module spending half its hours at 25 °C and half at 65 °C averages 45 °C, but its true damage rate is the chord midpoint (AF ≈ 8.4, equivalent to a constant 54.9 °C), nearly double the AF ≈ 4.3 of a module actually held at 45 °C. That is why Teq is computed from all 8,760 hourly temperatures instead of their mean — the hot afternoons do disproportionate damage that a cool night cannot repay. Original figure.
(12.1) Arrhenius equivalent temperature
Teq=−Ea/kln⁡ ⁣(1N∑i=1Nexp⁡(−Ea/(k Ti)))−273.15 [∘C]{\htmlData{sym=T_eq}{T_{\text{eq}}}} = \dfrac{-{\htmlData{sym=E_a}{E_a}}/{\htmlData{sym=kb}{k}}}{\ln\!\Big(\tfrac{1}{{\htmlData{sym=N_hrs}{N}}}\textstyle\sum_{{\htmlData{sym=i_hr}{i}}=1}^{{\htmlData{sym=N_hrs}{N}}}\exp\big({-{\htmlData{sym=E_a}{E_a}}}/({\htmlData{sym=kb}{k}}\,{\htmlData{sym=T_i}{T_i}})\big)\Big)} - 273.15\ [^\circ\text{C}]
(12.2) Acceleration factor (Jordan–Kurtz 2013)
AF=exp⁡ ⁣[Eak(1Tref−1Teq)]{\htmlData{sym=AF}{\text{AF}}} = \exp\!\left[\dfrac{{\htmlData{sym=E_a}{E_a}}}{{\htmlData{sym=kb}{k}}}\left(\dfrac{1}{{\htmlData{sym=T_ref}{T_{\text{ref}}}}} - \dfrac{1}{{\htmlData{sym=T_eq}{T_{\text{eq}}}}}\right)\right]
(12.3) Climate-adjusted degradation rate
rate=rbase ⋅ AF [%/yr]{\htmlData{sym=ratef}{\text{rate}}} = {\htmlData{sym=r_base_v}{r_{\text{base}}}}\,\cdot\,{\htmlData{sym=AF}{\text{AF}}}\ [\%/\text{yr}]
(12.4) 25-year projection
Yn=Y0 (1−LID) (1−rate) n−1 [kWh/m2]{\htmlData{sym=Y_n}{Y_n}} = {\htmlData{sym=Y_0}{Y_0}}\,(1 - {\htmlData{sym=LID}{\text{LID}}})\,(1 - {\htmlData{sym=ratef}{\text{rate}}})^{\,{\htmlData{sym=n_year}{n}}-1}\ [\text{kWh/m}^2]
SymbolMeaningUnitsValue / source
TeqArrhenius equivalent temperature — the constant temperature causing the same total degradation as the real hourly historyK (shown °C)−(Ea/k) / ln(mean(exp(−Ea/(k·Ti)))) − 273.15; always above the arithmetic mean because hot hours are exponentially up-weighted
TiHourly cell temperature from §8Kclipped to [230, 420] K to prevent numerical overflow
NHourly samples in the year's history—8,760
EaActivation energy for overall c-Si power losseV0.6 — PV reliability literature
kBoltzmann constanteV/K8.617×10⁻⁵
TrefReference temperature (25 °C)K298.15
AFAcceleration factor vs the 25 °C reference—exp[(Ea/k)(1/Tref − 1/Teq)]; rate = rbase × AF
rbaseBaseline degradation rate at 25 °C%/yrwarranted rate → per-tech median → 0.5 default (NREL x-Si median)
LIDFirst-year degradation (LID plus other year-1 effects)%datasheet figure → per-tech warranty-class ceiling → 2 default — one-time step, year 1 only
YnYield in year n of 25kWh/m²Y₀ (1 − LID)(1 − rate)n−1; lifetime = Σ over 25 years
Y0Undegraded annual yield entering the projectionkWh/m²the §14 annual yield

Wind-corrected Tcell (Faiman, §8) feeds directly into Teq — a windier site degrades more slowly over 25 years as well as producing more energy annually. This irreversible decline is distinct from the instantaneous thermal derating in §9: that one reverses when the panel cools; this one does not.

13. Perfect Cooling Scenario▾

Theoretical maximum: identical to two-axis tracking but cell temperature is capped at 25 °C at all times (Tcell = min(Tcell,Faiman, 25.0 °C)). Physically achievable with active liquid cooling or forced-air heat exchangers. Represents the ceiling for active thermal management and also carries the lowest 25-year degradation rate of all configurations — its cell never exceeds standard test conditions, so the Arrhenius model runs at the 0.5%/yr baseline (its own 25-year profile is computed alongside the others).

Concept — one cap, two payoffs

Clamping the cell at 25 °C pays out through two independent mechanisms, which is why the scenario exists as an engineering ceiling rather than a curiosity. First, the instantaneous one (§9): no temperature excess, no derating — every hour runs at rated conversion efficiency. Second, the cumulative one (§12): the Arrhenius acceleration is computed from the cell-temperature history, and a capped history never exceeds the reference point, so degradation runs at the 0.5%/yr baseline — the slowest of any configuration. The gap between this line and the standard two-axis line on the 25-year chart is therefore the total prize active thermal management could ever claim at this site: a real cooling system (with its own pumping energy costs, not modelled) can only approach it.

(13.1) Active-cooling ceiling
Tcell=min⁡(Tcell,Faiman, 25.0 ∘C) [∘C]{\htmlData{sym=T_cell}{T_{\text{cell}}}} = \min({\htmlData{sym=T_cellF}{T_{\text{cell,Faiman}}}},\ 25.0\ {^\circ}\text{C})\ [^\circ\text{C}]
14. Output Metrics▾

Hourly power (§9) is summed to annual energy, and the capacity factor (CF) expresses that energy as a fraction of what the module would produce running at its full rated output every hour of the year:

Concept — capacity factor — how full is the year?

Annual yield in kWh/m² is honest but hard to compare across technologies; the capacity factor rescales it into a fraction anyone can read: of the energy this module would deliver if it ran at its full rated output every single hour of the year, what share did it actually deliver? The 87.6 in the denominator is just that hypothetical year in disguise — 8,760 hours × 1 kW/m² of standard test irradiance, divided by 1,000 — scaled by the module efficiency η. Nights alone cap CF near 50%; add mornings, evenings, winter and cloud and even a desert tracker lands near 30%. That is not underperformance — it is the honest arithmetic of a planet that rotates.

(14.1) Annual energy yield
Eannual=11000∑t=18760P(t)  [kWh/m2]{\htmlData{sym=E_annual}{E_{\text{annual}}}} = \frac{1}{1000}\sum_{{\htmlData{sym=t_time}{t}}=1}^{8760} {\htmlData{sym=Ppow}{P}}({\htmlData{sym=t_time}{t}})\ \ \text{[kWh/m}^2\text{]}
(14.2) Capacity factor
CF=Eannualη⋅87.6  [%]{\htmlData{sym=CFv}{\text{CF}}} = \frac{{\htmlData{sym=E_annual}{E_{\text{annual}}}}}{{\htmlData{sym=eta}{\eta}} \cdot 87.6}\ \ [\%]
(14.3) Tracking gain
gain=E2ax−EftEft×100  [%]{\htmlData{sym=gainf}{\text{gain}}} = \frac{{\htmlData{sym=E_2ax}{E_{\text{2ax}}}} - {\htmlData{sym=E_ft}{E_{\text{ft}}}}}{{\htmlData{sym=E_ft}{E_{\text{ft}}}}} \times 100\ \ [\%]
SymbolMeaningUnitsValue / source
EannualAnnual energy yield per unit areakWh/m²sum of hourly W/m² ÷ 1000
CFCapacity factor — energy delivered vs rated-output-always%denominator η·87.6 = rated STC energy: 8,760 h × 1 kW/m² × η ÷ 1000
E2axTwo-axis tracking annual yieldkWh/m²computed every run
EftFixed-tilt reference annual yieldkWh/m²computed every run
gainTracking gain of two-axis over the fixed-tilt reference%(13.3)

Typical two-axis CF: ~28–32% desert, ~20–25% mid-latitude, ~10–14% temperate/cloudy. The dashboard also reports peak plane-of-array irradiance (the maximum hourly W/m² encountered in the year). The η in the denominator is the resolved module efficiency — your module's derived η when parameters are set, 0.20 otherwise — so CF stays η-invariant (a more efficient module yields more and is rated higher, in the same proportion). With module parameters set, the dashboard additionally reports single-module energy = E × module area (kWh/yr — one module; array totals arrive with the array design tools).

15. Climate Data Representativeness — ENSO & PDO▾

Every climate input comes from a multi-year average, and the averaging window determines how representative it is of a typical year. The El Niño–Southern Oscillation (ENSO) is the dominant source of year-to-year climate variability globally, shifting rainfall, wind, and temperature every 3–7 years; the Pacific Decadal Oscillation (PDO) modulates ENSO strength on a ~20–30 year cycle.

Concept — averaging over a see-saw

A "climatological normal" sounds objective, but it is only as representative as the years inside the window. When the underlying quantity oscillates, a window shorter than a full cycle can land entirely on one phase — and its average faithfully reports that phase, not the climate (Fig. 15.1). That is a sampling bias: no measurement is wrong, and more data of the same years would not fix it; only spanning whole cycles does. The rule of thumb applied below is to size each variable's window in cycles of its dominant driver, not in years — which is why rainfall gets 20 years (≈4 ENSO cycles) while aerosols, which do not follow ENSO, get by with 5.

El NiñoLa Niña20002005201020152020triple La Niña2018–2022 — 5 yr, lands on one phase: biased2001–2020 — ≈4 full cycles: phases cancel
Fig. 15.1 — Averaging over a see-saw (schematic oscillation — not a measured index). Climate swings between El Niño and La Niña every 3–7 years, and a "climatological average" is only as representative as the window it averages. The original 5-year window (2018–2022) sat squarely on the triple La Niña — its rainfall mean was wet-biased at every test site (Lima +125%) — while a ~20-year window spans several full cycles, letting the wet and dry phases cancel. Hence the per-variable window policy in the table below. Original figure.

The original ERA5 2018–2022 window coincided with a triple La Niña — quantified comparisons at five sites showed wet bias at every location (Lima +125%, Riyadh +95%, Bangkok +14%), causing the soiling model to overestimate natural cleaning and underestimate dust accumulation in typical or El Niño years. Each input's window is now chosen deliberately:

InputSourceWindowENSO cyclesRationale
RainfallNASA POWER MERRA-22001–2020 (20 yr)~4Avoids triple La Niña; strongest ENSO signal
Wind speedNASA POWER MERRA-22001–2020 (20 yr)~4Same call as rainfall; consistent window
Surface albedo (seasonal mode)NASA POWER MERRA-22001–2020 (20 yr)Low sensitivitySame call as rainfall; a 20-yr mean shape smooths drought/wet-year extremes
SnowfallOpen-Meteo ERA52004–2022 (19 yr)~3NASA POWER snow depth is a different physical quantity (state, not flux)
TemperaturePVGIS TMY2005–2020 (16 yr) composite~3TMY month selection is distribution-matched, not single-year
IrradiancePVGIS TMY2005–2020 (16 yr) composite~3
AOD (dust)NASA POWER MERRA-22018–2022Low sensitivityAerosol cycles are not ENSO-driven; 5-year window acceptable
16. Interannual Variability & P90 Yield Confidence▾

The PVGIS TMY provides a single typical-year yield — the P50, exceeded in half of all years. P90 (exceeded in 9 years out of 10) and P75 bounds are derived by combining that P50 with a precomputed global coefficient of variation (CV) from 39 years of historical simulation: NASA POWER 1984–2022, 256,573 land cells at 0.5° resolution, 4 soiling-multiplier variants per cell.

Concept — one number for a typical year, three for an honest one

A single yield estimate hides a question every financier asks: how bad can a bad year get? Annual yields scatter around the typical value — cloudier summers, dustier winters — and the spread is summarised by the coefficient of variation (CV), the standard deviation as a fraction of the mean. The downside bounds then sit at fixed depths into the left shoulder of that spread (Fig. 16.1): P75 at 0.674 standard deviations below the median, P90 at 1.28. At a typical CV of 8%, P90 = P50 × (1 − 1.28 × 0.08) ≈ 90% of P50 — the yield nine years in ten will beat, and the number debt service gets sized against. The anchor (P50) comes from the live PVGIS simulation; only the spread comes from the precomputed 39-year grid.

annual yield → (spread: one σ = CV × P50)P50typical yearP753 in 4 exceedP909 in 10 exceedthe 1-in-10 bad years
Fig. 16.1 — Where the downside bounds sit on the spread of annual yields (assumed ≈ normal about the median). P50 is the typical year — half of all years beat it. P75 and P90 slide down the left shoulder by the standard-normal quantiles the model uses: 0.674 and 1.28 standard deviations, where one σ = CV × P50. Reading P90: nine years in ten will yield more than this — the number a lender sizes debt service against. At a typical CV of 8%, P90 = P50 × (1 − 1.28 × 0.08) ≈ 90% of P50. Original figure.
(16.1) Combine variability (RSS)
cvtotal=cvclimate2+cvsnow2{\htmlData{sym=cv_total}{\text{cv}_{\text{total}}}} = \sqrt{{\htmlData{sym=cv_climate}{\text{cv}_{\text{climate}}}}^2 + {\htmlData{sym=cv_snow}{\text{cv}_{\text{snow}}}}^2}
(16.2) Downside bounds
P90=P50 ⋅ (1−1.28 ⋅ CV) [kWh/m2]P75=P50 ⋅ (1−0.674 ⋅ CV) [kWh/m2]\begin{aligned} {\htmlData{sym=P90v}{P90}} &= {\htmlData{sym=P50v}{P50}}\,\cdot\,(1 - 1.28\,\cdot\,{\htmlData{sym=CV}{\text{CV}}})\ [\text{kWh/m}^2] \\ {\htmlData{sym=P75v}{P75}} &= {\htmlData{sym=P50v}{P50}}\,\cdot\,(1 - 0.674\,\cdot\,{\htmlData{sym=CV}{\text{CV}}})\ [\text{kWh/m}^2] \end{aligned}
SymbolMeaningUnitsValue / source
P50Median-year yield — the anchorkWh/m²live PVGIS TMY simulation (absolute scale)
CVCoefficient of variation of annual yields = σ/μ—precomputed grid; sample standard deviation (ddof = 1) over the 39 simulated years
zStandard-normal quantiles—1.28 → 10th percentile (P90); 0.674 → 25th percentile (P75)
cvclimateClimate variability component—39-yr NASA POWER joint simulation
cvsnowSnow variability component—19-yr ERA5 snow record (2004–2022), snow-affected cells only
grid cellNearest 0.5° cell centre°⌊x/0.5⌋·0.5 + 0.25; soiling multiplier snapped to nearest of ×1/×3/×8/×15; no P90 shown for ocean/absent cells

Grid inputs (one NASA POWER call per cell, all 39 years): ALLSKY_SFC_SW_DWN (GHI, kWh/m²/day — available from 1984, which fixes the window start), T2M (°C), PRECTOTCORR (mm/day), WS10M (m/s), and AOD_55 (reliable from 2001; earlier years use the cell's own 2001–2022 monthly means). Each year runs the Faiman thermal model, temperature derating, and the full soiling model — then CV is taken over the 39 annual yields.

Why joint simulation matters: each year varies irradiance, temperature, rainfall, wind, and AOD simultaneously — capturing compensating real-world correlations (La Niña years: more rain → cleaner panels, but less sun → cooler cells). Industry-standard tools typically vary irradiance only. Treating inputs as independent worst cases would overstate variability; the joint approach is less pessimistic and more physically defensible. PVGIS provides the absolute scale (P50); NASA POWER contributes only the relative spread — so NASA POWER's absolute irradiance biases cancel in the ratio. P90 is not computed on the clear-sky fallback path (no meaningful P50 anchor). The dashboard's fixed-tilt reference P90/P75 bars apply the same CV to the fixed-tilt P50: at one location both systems face the same climate variability, and CV is a relative measure (the same approximation as the GHI proxy — see Statistical Methods).

Snow variability (RSS correction): NASA POWER has no snowfall variable, so the main grid omits snow variability. For the 42,513 grid cells with historical snow, ERA5 2004–2022 records give a separate snow CV (year-by-year snow loss normalised by the cell's mean yield), combined in quadrature — root-sum-square (RSS): cvtotal = √(cvclimate² + cvsnow²). RSS is valid because the drivers are independent: snow variability follows the Arctic / North Atlantic Oscillation, largely orthogonal to the ENSO / PDO signals driving climate CV — which also means the two CVs need not share the same year windows.

cvclimate = 8% — ENSO / PDO drivencvsnow= 4% — AO / NAOcvtotal = √(8² + 4²) ≈ 8.9%
Fig. 16.3 — Why independent variabilities combine as a root-sum-square: they add like perpendicular directions — Pythagoras for randomness. When two sources of year-to-year spread are uncorrelated (snow rides the Arctic/North Atlantic Oscillation; the climate term rides ENSO/PDO), their variances add, so the combined CV is the hypotenuse, not the sum: √(8² + 4²) ≈ 8.9%, noticeably less than the 12% naive addition would give. Correlated sources would need the sum; independence earns the discount. Original figure.

What is a Typical Meteorological Year (TMY)? A TMY is a synthetic dataset constructed by independently selecting one real historical month for each of the 12 calendar months from a multi-year satellite record (PVGIS uses the 16 years 2005–2020). Selection uses the Finkelstein-Schafer (FS) statistic: for each candidate month, the empirical cumulative distribution function (CDF) of each climate variable — GHI, DNI, diffuse irradiance, temperature, wind speed — is compared against the long-run CDF for that calendar month. The FS score measures the area between those two CDFs (similar in concept to a Kolmogorov-Smirnov test). Scores are weighted across variables — GHI carries roughly 50% of the weight — and the month with the lowest weighted FS score is selected: the month whose full statistical distribution most closely matches the long-run distribution for that slot.

25%50%75%100%daily value of one climate variable (e.g. GHI) →fraction of days at or below that value (CDF)long-run CDF — 16 years pooledone candidate monthFS score = the area between the curves
Fig. 16.2 — How one TMY month is selected (schematic curves). For each candidate — the 16 real Junes, say — the cumulative distribution function (CDF) of each climate variable is compared against the long-run CDF pooled across all 16 years. The Finkelstein–Schafer score is the area between the two curves (the tinted gap): a candidate matching the long-run distribution — median, spread, cloudy-day frequency, all of it — scores near zero. Scores are weighted across variables (irradiance ≈ half the weight) and the lowest total wins the slot. Matching whole distributions, not means, is why a TMY behaves like a typical year rather than an averaged one. Original figure.

The TMY is not the average year. It matches the full shape of the distribution — including spread — not just the mean. Two months with identical means but different variances produce different FS scores. The assembled year is not a year that ever actually occurred: it is a patchwork of up to 16 different real years, one month at a time. It sits close to the long-run median by construction, but the selection is discrete — only 16 real months to choose from per slot, not a continuous average.

Monthly chart discontinuities: Because each month is selected independently, June might come from a cloudy year while May comes from a clear one — both were the most distribution-representative months for their respective slots, but when concatenated the transition looks like a step. At locations where interannual irradiance variability is low, all 12 months come from years that look similar so the stitching is invisible. At locations with high year-to-year variability, the same calendar month can vary 2–3× between years, and the concatenation produces visible steps or inversions in the monthly profile (e.g. May appearing higher than June). These are not modelling errors: they reflect the discrete, independent nature of TMY construction applied to a genuinely erratic climate.

17. Bifacial Rear Side▾

A bifacial module also converts light arriving at its back — mostly sunlight the ground reflects. Entering a bifaciality on the module form switches this model on for both configurations: the same module, each earning its own rear bonus from its own geometry. The rear contribution joins the front as an effective irradiance that drives the whole downstream chain — cell temperature (§8), temperature and low-light response (§9), degradation (§12) and capacity factor (§14).

Concept — why the fixed panel wins the percentage — the tracker stares at its own shadow

Set a bifaciality and the results show the fixed panel earning the larger rear-side gain — which looks backwards for a machine built to chase the light. The resolution (Fig 17.1): perfect front-side aim is worth nothing to the rear, and less than nothing nearby. Pointing the front at the sun points the rear at the module's own shadow. A tracker's rear is always exactly anti-sun, so direct beam never reaches it (eq 17.3's rear-beam term is identically zero for the tracker), and its shadow always falls on precisely the ground the rear faces. High sun is the worst case: the tracker lies nearly flat (Σ = 90° − β), looking almost straight down at that shadow — at its full width, because squarely facing the sun also casts the largest shadow a module can.

The fixed panel aims imperfectly, and its rear profits from the sloppiness. Standing at a constant tilt, it looks across the ground rather than down at it, keeps a permanent wedge of sky diffuse in frame, and even collects direct beam on the rear when the sun passes behind its plane (summer mornings and evenings at higher latitudes). The two rear sides therefore end up collecting roughly the same light over the year — the tracker a touch ahead over dark ground, a touch behind over bright ground, where the beam shadow it casts on its own view costs the most.

The denominator settles it. The reported gain is relative — each configuration's rear bonus divided by its own monofacial yield — and the tracker's denominator is the largest any geometry can produce (§5: no cosine loss, ever). A similar rear bonus over a much larger base is a smaller percentage. None of this dethrones the tracker: its total yield keeps the full tracking advantage and the rear bonus lifts it further — the fixed panel wins the percentage, not the year.

frontrearown shadowa wedge of skyfixed tilt — rear looks across the grounda steady wide view: sky, shadow, and ground beyond itfrontrearrear always exactly anti-sun— direct beam never reaches itown shadow — the full width of the moduletwo-axis — near-flat at high sun (Σ = 90° − β)rear looks straight down — at its own full-width shadow
Fig. 17.1 — Two rear-side worlds. The tracker (right) aims its front perfectly — which points the rear at its own shadow: an exactly anti-sun face never receives direct beam, and near noon the module lies almost flat over a full-width shadow of itself (facing the sun squarely also casts the largest shadow a module can). The fixed panel (left) aims imperfectly and its rear profits: it looks across the ground rather than down at it and keeps a wedge of sky in frame — and it is drawn here at its sun-facing hour; through the rest of the day its shadow swings sideways, out of the rear's view. Over the year the two rear sides collect roughly the same light — but the tracker's front collects far more (§5, no cosine loss), so the same-sized rear bonus reads as a smaller percentage of it. Original figure.

Ground clearance is the rear side's lever — the module dims the very ground it depends on, and elevation buys that ground back (Fig 17.2). Sitting low, the rear's near field is the module's own shadow plus ground it has starved of sky light. Raised, sunlight reaches the ground beneath, the shadow is pushed out into a much wider visible sweep, and the rear brightens — steeply over roughly the first metre, then levelling off (Yusufoglu et al., via Sun et al. 2018), which is why the clearance input defaults inside the recommended 0.5–1.5 m band. The same clearance is applied to both configurations, so the comparison isolates the tracking geometry. In the equations below, E enters through the view-factor geometry of eq 17.1: the module's edges sit at heights E and E + ℓ sin Σ above the ground, and that placement sets every string length in the crossed-strings construction — hence every view factor — as well as where the beam shadow lands and how much sky each ground segment still sees.

rearown shadowsitting low — E smallthe near field: its own shadow, dim sky-starved groundElight gets underneathshadow — pushed away from the rear's near fieldraised — about a metrelight floods beneath; the shadow loses its grip on the view
Fig. 17.2 — Clearance E buys the rear its view. Sitting low (left), the rear's near field is the module's own shadow plus ground the module has starved of sky light. Raised (right), sunlight reaches the ground beneath and the shadow is pushed out into a much wider visible sweep — the rear brightens, steeply over roughly the first metre of clearance, then levelling off (Yusufoglu et al., via Sun et al. 2018). The same clearance is applied to both configurations, so the two-axis vs fixed-tilt comparison isolates tracking geometry. Original figure.

How the ground's light is booked — four strings per segment: the strip of ground the rear can see is split into 120 segments of width wi, spanning ±15 slant lengths either side; everything beyond is a single analytic far-field term. For each segment, each hour, the model needs its view factor Fi→p — of all the light leaving segment i, the fraction that arrives at the module. Hottel's crossed-strings construction (eq 17.1, Fig 17.3) reads that fraction straight off the geometry: stretch four imaginary strings from the segment's two endpoints to the module's two edges. Two of them cross each other — the diagonals of the quadrilateral formed by segment and module — and two do not (its free sides). Their summed lengths are dcross and duncross, and the view factor is the crossed pair's excess over the uncrossed, per unit of segment width: Fi→p = (dcross − duncross)/(2wi). The excess measures how much of the segment's outlook the module fills — a distant or edge-on segment has near-equal string lengths and a view factor near zero; ground the module leans over has a large one. This is where clearance and slant length act: the module's edges sit at heights E and E + ℓ sin Σ, and every string is measured to those two points.

The same exchange is then re-booked from the module's side by reciprocity: Fp→i = Fi→p · wi/ℓ — a view factor counts a shared exchange relative to the surface doing the looking, so swapping viewpoints just swaps the two widths. What segment i has to offer, Ii, is its unshaded direct beam — the fraction 1 − shadedi left uncovered by the module's beam shadow that hour — plus the sky diffuse the module has not masked from it, Fi→sky = 1 − Fi→p (in the 2-D cross-section the module is the only obstruction). Each segment's reflected light ρ·Ii then reaches the rear with weight Fp→i; summed over all 120 segments plus the far field, that is the mean ground irradiance Īgnd of eq 17.2, which enters the rear plane-of-array total (eq 17.3) through the §6.3 rear-plane angular factor.

rearEℓΣcrosseduncrossed (dashed)wi — segment ifour strings tie the segment to the moduleFi→p = (dcross − duncross) / 2wi — eq 17.1beam shadowshadediopen skymaskedIi: beam where unshaded, sky where unmaskedρ·Ii·Fp→i joins Īgnd — eq 17.2
Fig. 17.3 — The rear side's ground bookkeeping, one segment at a time. Left — the four strings of eq 17.1: from the segment's two endpoints to the module's two edges (which sit at heights E and E + ℓ sin Σ — that is how clearance and slant length enter). The solid pair cross — they are the diagonals of the segment–module quadrilateral; the dashed pair are its free sides and do not. The view factor Fi→p = (dcross − duncross)/(2wi) is the crossed pair's excess length over the uncrossed, per unit of segment width: a distant or edge-on segment has near-equal strings and a view factor near zero; ground the module leans over has a large excess. Reciprocity re-books the same exchange from the module's side: Fp→i = Fi→p · wi/ℓ. Right — what segment i carries (eq 17.2): direct beam on its unshaded fraction (1 − shadedi), plus sky diffuse on the share of sky the module leaves it (Fi→sky = 1 − Fi→p); its reflected light ρ·Ii then climbs to the rear with weight Fp→i, summing to Īgnd. Original figure.

A note on the symbol: datasheets and IEC TS 60904-1-2 print bifaciality as a bare φ. This model writes it bbif — φ is reserved exclusively for the azimuth family (φs solar, φp panel), the same reservation that writes latitude as L, so a reader never meets one letter carrying two meanings inside an equation.

(17.1) Panel–ground view factors (crossed strings)
Fi→p=dcross−duncross2 wi,Fp→i=Fi→p wiℓ{\htmlData{sym=F_ip}{F_{i\to p}}} = \frac{{\htmlData{sym=d_cross}{d_{\text{cross}}}} - {\htmlData{sym=d_uncross}{d_{\text{uncross}}}}}{2\,{\htmlData{sym=w_i}{w_i}}}, \qquad {\htmlData{sym=F_pi}{F_{p\to i}}} = {\htmlData{sym=F_ip}{F_{i\to p}}}\,\frac{{\htmlData{sym=w_i}{w_i}}}{{\htmlData{sym=ell}{\ell}}}
(17.2) Visible-ground irradiance (self-shading, both categories)
Ii=DNI sin⁡β (1−shadedi)+Id Fi→sky    ⇒    I‾gnd=∑iIi Fp→i/1+cos⁡Σ2{\htmlData{sym=I_i}{I_i}} = {\htmlData{sym=DNI}{\text{DNI}}}\,\sin{\htmlData{sym=beta}{\beta}}\,(1-{\htmlData{sym=shaded_i}{\text{shaded}_i}}) + {\htmlData{sym=I_d}{I_d}}\,{\htmlData{sym=F_isky}{F_{i\to sky}}} \;\;\Rightarrow\;\; {\htmlData{sym=I_gnd}{\overline{I}_{\text{gnd}}}} = \sum_{\htmlData{sym=i_strip}{i}} {\htmlData{sym=I_i}{I_i}}\,{\htmlData{sym=F_pi}{F_{p\to i}}} \Big/ \tfrac{1+\cos{\htmlData{sym=Sigma}{\Sigma}}}{2}
(17.3) Rear plane-of-array irradiance
Srear=Id (1−Ai) 1−cos⁡Σ2 Fd(180∘ ⁣− ⁣Σ)+DNI max⁡(0,−cos⁡θ) IAM+ρ I‾gnd 1+cos⁡Σ2 Fa(180∘ ⁣− ⁣Σ)\begin{aligned} {\htmlData{sym=S_rear}{S_{\text{rear}}}} &= {\htmlData{sym=I_d}{I_d}}\,(1-{\htmlData{sym=A_i}{A_i}})\,\tfrac{1-\cos{\htmlData{sym=Sigma}{\Sigma}}}{2}\,{\htmlData{sym=F_d}{F_d}}(180^\circ\!-\!{\htmlData{sym=Sigma}{\Sigma}}) + {\htmlData{sym=DNI}{\text{DNI}}}\,\max(0,-\cos{\htmlData{sym=theta}{\theta}})\,{\htmlData{sym=IAM}{\text{IAM}}} \\ &\quad + {\htmlData{sym=rho}{\rho}}\,{\htmlData{sym=I_gnd}{\overline{I}_{\text{gnd}}}}\,\tfrac{1+\cos{\htmlData{sym=Sigma}{\Sigma}}}{2}\,{\htmlData{sym=F_a}{F_a}}(180^\circ\!-\!{\htmlData{sym=Sigma}{\Sigma}}) \end{aligned}
(17.4) Effective irradiance (IEC TS 60904-1-2)
Se=S+bbif (1−fstruct) Srear [W/m2]{\htmlData{sym=S_e}{S_e}} = {\htmlData{sym=Sirr}{S}} + {\htmlData{sym=b_bif}{b_{\text{bif}}}}\,(1 - {\htmlData{sym=f_struct}{f_{\text{struct}}}})\,{\htmlData{sym=S_rear}{S_{\text{rear}}}}\ [\text{W/m}^2]
SymbolMeaningUnitsValue / source
bbifBifaciality — rear-side efficiency as a fraction of front—datasheet (printed there as a bare φ; written b — see the notation note above); the lesser of the Isc and Pmax rear/front ratios at STC (IEC TS 60904-1-2); typical p-PERC 0.60–0.70, HJT above 0.90
fstructRear structure shading factor—default 0.05 — neither mount carries a torque tube, and open racking sits at the low end; published spread 2–20% by structure (IEA PVPS Task 13, 2021, §7); user 0–0.30
SrearRear plane-of-array irradianceW/m²eq 17.3 — isotropic-remainder sky diffuse (Id·(1 − Ai), §5 — circumsolar rides the front beam path) + beam (fixed tilt only) + ground reflection
ĪgndMean visible-ground irradianceW/m²eq 17.2 — view-factor-weighted over 120 ground segments (±15 slant lengths); analytic far field beyond
Fi→pSegment→module view factor—eq 17.1, Hottel's crossed strings: (dcross − duncross)/(2wi) — of all the light leaving segment i, the fraction that arrives at the module (Fig 17.3)
dcrossCrossed string lengths, summedmthe two strings that cross — the diagonals of the quadrilateral formed by the segment's endpoints and the module's edges (Fig 17.3); their excess over the uncrossed pair is the exchange
duncrossUncrossed string lengths, summedmthe two strings that don't cross — the quadrilateral's free sides; equal to the crossed pair for a segment the module barely appears to, so the view factor vanishes with distance
wiGround segment widthm120 segments spanning ±15 slant lengths either side of the module; beyond them one analytic far-field term (unshaded, full sky view) closes the rear's total ground view factor exactly to (1 + cos Σ)/2
Fp→iModule→segment view factor—reciprocity: Fi→p · wi/ℓ — the same exchange booked from the module's side; the weight carrying each segment's ρ·Ii into Īgnd
IiGround-segment irradianceW/m²eq 17.2 — unshaded beam, DNI sin β (1 − shadedi), plus the sky diffuse the module does not mask, Id Fi→sky
shadediSegment beam-shadow fraction—fraction of segment i inside the module's beam shadow that hour — each module edge projected along the sun direction onto the ground; for the tracker the shadow falls exactly where the rear looks (Fig 17.1)
Fi→skySegment→sky view factor—1 − Fi→p — the module is the only obstruction in the 2-D cross-section; whatever slice of sky the module doesn't block, the segment receives as diffuse light
ρSite ground albedo—the §2 site surface input — the dominant driver of the rear bonus; in months passing the §11 snow gate every surface sees it lifted to the snow value
EGround clearance (lower-edge height)mmodule form; default 1.0 m — inside the recommended 0.5–1.5 m band (Yusufoglu et al., via Sun et al. 2018). Enters eq 17.1's crossed-strings geometry as the module edge heights E and E + ℓ sin Σ
ℓModule slant length (cross-section)mfixed at 2.0 m — named simplification; the rear gain depends on the clearance-to-slant ratio only weakly over the 1.7–2.3 m module band

Self-shading — why the model matters: the module dims the very ground its rear looks at, in two ways. Its beam shadow (each edge projected along the sun direction) removes direct sun from the segments it covers — and for the tracker the shadow falls exactly where the rear faces, since the rear is always anti-sun. Separately, the module masks part of each ground segment's view of the sky, dimming its diffuse light. Both effects are computed per segment (crossed-strings view factors), following the standalone-module framework of Sun, Khan, Deline & Alam (2018) with ground-segment configuration factors after Marion et al. (2017). Ground beyond the modelled extent is treated analytically (unshaded, full sky view), so the rear's total ground view factor closes exactly to (1 + cos Σ)/2. Skipping self-shading would systematically flatter the tracker — the opposite of this comparison's job.

Rear angular losses reuse the §6.3 Martín–Ruiz integrated factors evaluated at the rear plane (tilted 180° − Σ) — the same physical role as the angle-of-incidence corrections in Marion et al., using this model's existing machinery. The model class validates well: Marion et al. report total-irradiance mean bias within ±2.4% against field measurements, and Sun et al. reproduce published gains within about ±5 percentage points; this implementation lands rear-to-front ratios of roughly 10–16% over grass (measured ≈ 10%) and yield gains inside the published bands, pinned by tests.

Snow-brightened ground: in months whose snowfall passes the §11 gate, the effective ground albedo rises to the snow value for every surface — both configurations' fronts (through the ground-reflected term of §2), both rears, and the §7 tilt sweep — so the whole model agrees about the brightness of the same patch of ground, and the hero yields carry the effect directly. The §11 albedo-bonus figure remains as an attribution estimate of the fixed-tilt front's share, not an addition on top.

Soiling stays a front-side affair: dust settles on the upward glass, so the §10 soiling model differences front-only monofacial yields against its front-side αr table — the rear bonus can neither soil nor be counted as a soiling loss — while the reported percentages divide by the configuration's full (bifacial) yield.

Deliberately not modelled (each small, named, and conservative in direction): rear-side soiling; the sub-0.5% electrical mismatch from rear non-uniformity on a single module (Deline et al., via IEA PVPS Task 13, 2021); and the third dimension. The 2-D cross-section is equivalent to an infinitely long row of modules, which slightly understates the unshaded ground a real single module sees past its own ends — a conservative error worth a few per cent of the rear term, i.e. of order a few tenths of a per cent of total yield, and it largely cancels in the two-axis vs fixed-tilt comparison because both configurations share the geometry. Full 3-D treatments exist — finite-element view-factor models (e.g. the DUET model of Tonita et al. 2024) and ray tracing (NREL's bifacial_radiance) — but a 3-D view factor is no longer a strings construction: it requires per-patch solid-angle integrals over a two-dimensional ground grid, re-evaluated hourly for a tracker whose geometry never repeats. That cost would buy precision below the uncertainty of the albedo input itself, which dominates any bifacial estimate (IEA PVPS Task 13, 2021): spectral albedo effects alone move the rear-side spectral mismatch by up to ±13% and the total by up to 3% (Tonita et al. 2024) — the §2 spectral factor now corrects the static part of that, but the albedo level uncertainty remains. The industry's bankable tools make the same choice — PVsyst's bifacial model and NREL's bifacialvf (the Marion et al. 2017 model this implementation follows) are 2-D view-factor models. The same clearance is applied to both configurations so the comparison isolates tracking geometry — real two-axis pedestals often sit higher, an unmodelled upside for the tracker.

Energy Storage · Distribution / Consumption

18. System & Storage — Wiring, Conversion, Battery▾

Beyond the panels, the model carries the array’s hourly energy through the REAL electrical chain — string wiring, conversion, and a battery modelled as a physical device — with every stage anchored to what component datasheets actually print, never to invented efficiencies. The order of ideas is deliberate: battery capability comes before any load profile. What a pack can physically absorb and deliver sets the design constraints; a consumption profile is then an analysis input checked against that capability (its shortfall directly sizes backup generation for off-grid designs).

String wiring. Panels wired in series must carry one shared current, so a shaded panel drags its string. Each hour, every panel’s own current–voltage curve (§17 electrical model) is evaluated at its own shaded light level, summed in series per string, and the true string power is compared with what per-panel electronics would collect:

Hours where a string’s panels see equal light couple losslessly by construction (the maximum power of identical series curves is additive), so only genuinely mismatched hours are solved. A constant manufacturing-dispersion factor (2% crystalline / 1% thin-film; PVsyst array-mismatch documentation, 2026) covers panel-to-panel spread, and cold-morning / hot-afternoon panel voltages from the site’s own simulated hours are checked against each device’s printed input window. A PWM (pulse-width modulation) charge controller changes the operating point entirely — the string is pinned at the battery bus instead of tracking its own best voltage:

panel voltage (V)panel current (A)maximum power pointpinned at the busbus ≈ 27 V41 Vlost ≈ 28%(the strip between rectangles)
Fig. 18.4 — Why PWM coupling costs harvest: the panel can only deliver power at ONE point on its current–voltage curve, and a PWM controller pins that point at the battery bus voltage (here a 24 V-class bank, bus ≈ 27 V) instead of letting the panel sit at its maximum power point (≈ 41 V). The delivered power is the dashed rectangle under each operating point — voltage × current — and the missing strip between them is the PWM penalty (eq 18.2): here ≈ 28% at standard conditions, exactly the 5–30% class controller literature reports. Cold makes it worse (the maximum-power voltage rises AWAY from the bus); a bus above the curve's reach at the far right harvests nothing at all. Model-drawn single-diode curve (430 W module, standard conditions).

Worked example — what one panel loses on a 24 V bank

A 430 W module (best-power point 41 V × 10.0 A at standard conditions) wired through a PWM controller to a 24 V-class lithium bank: the bus sits near 27.4 V (midway between nominal 25.6 and absorption 29.2 V). Pinned there, the panel's own curve gives 10.9 A — 27.4 V × 10.9 A = 298 W against the 414 W available at the best point: a factor of 0.72 (eq 18.2, Fig. 18.4). The same panel on a 48 V bank cannot charge it at all — the bus exceeds what the curve can reach — while two in series restore the margin: string design IS the PWM decision.

(18.1) String coupling multiplier (per hour; shade mismatch emerges from the circuit)
mstr=∑stringsmax⁡[Istr⋅∑panelsV(Istr)]∑panelsPmp{\htmlData{sym=m_strv}{m_{\text{str}}}} = \dfrac{\sum_{\text{strings}} \max\big[ {\htmlData{sym=I_strv}{I_{\text{str}}}} \cdot \textstyle\sum_{\text{panels}} {\htmlData{sym=V_mod}{V}}({\htmlData{sym=I_strv}{I_{\text{str}}}}) \big]}{\sum_{\text{panels}} {\htmlData{sym=P_mp}{P_{\text{mp}}}}}
(18.2) PWM direct coupling — the string pinned at the battery bus
fpwm=Vbus  Istr(Vbus)∑Pmp{\htmlData{sym=f_pwm}{f_{\text{pwm}}}} = \dfrac{{\htmlData{sym=V_bus}{V_{\text{bus}}}}\; {\htmlData{sym=I_strv}{I_{\text{str}}}}({\htmlData{sym=V_bus}{V_{\text{bus}}}})}{\sum {\htmlData{sym=P_mp}{P_{\text{mp}}}}}

The curves are evaluated in the 25 °C frame with the datasheet temperature coefficients moving the operating point — the fitted curve’s own thermal slide is ~3× the datasheet value and never drives energy (the §9.3 doctrine). Cold worsens PWM (the panels’ best voltage rises AWAY from the bus), and a bus above what the string can reach draws exactly zero — one residential module cannot charge a 48 V bank, and the model says so.

Conversion. Inverter efficiency follows the PVWatts part-load form (NREL/TP-6A20-62641), anchored so the California Energy Commission-weighted average of the modelled curve equals the sheet’s printed weighted efficiency — an energy statement — or anchored to the printed efficiency curve itself when one is supplied:

(18.3) Part-load conversion efficiency (PVWatts form, NREL/TP-6A20-62641; weighted-point anchored)
ηinv(ζ)=ηnom0.9637 (−0.0162 ζ−0.0059/ζ+0.9858),ζ=Pdc/Pdc,0{\htmlData{sym=eta_inv}{\eta_{\text{inv}}}}({\htmlData{sym=zeta_ld}{\zeta}}) = \dfrac{{\htmlData{sym=eta_nomv}{\eta_{\text{nom}}}}}{0.9637}\,\big({-0.0162}\,{\htmlData{sym=zeta_ld}{\zeta}} - 0.0059/{\htmlData{sym=zeta_ld}{\zeta}} + 0.9858\big), \qquad {\htmlData{sym=zeta_ld}{\zeta}} = {\htmlData{sym=P_dcv}{P_{\text{dc}}}} / {\htmlData{sym=P_dc0}{P_{\text{dc,0}}}}

Output is capped at the alternating-current rating hour by hour (clipping), night standby draw is subtracted, and hot ambients plus thin high-altitude air lower the cap per the device’s printed derating chart or a cited class default (full output to a 45 °C knee then 2.5%/°C; altitude full to 2,000 m then 1%/100 m).

The battery is a circuit, not a percentage. Each chemistry (lithium iron phosphate; lithium nickel-manganese-cobalt, NMC; lead-acid) carries its own open-circuit voltage shape OCV(SOC), and the pack’s internal resistance is never guessed — it is derived from the datasheet’s own printed round-trip efficiency at its reference rate, so the model reproduces the sheet’s number exactly at the rating condition and every rate/temperature dependence of loss becomes emergent:

+−OCV(SOC)rest voltage — chemistry sets its shapeR0 — internal resistancefrom the printed round trip (eq 18.4)+−VtermI
Fig. 18.1 — The battery as a circuit. A rest-voltage source OCV(SOC) — whose shape the chemistry sets (Fig. 18.2) — sits behind one series resistance R0, and everything rate- and temperature-dependent flows from their product: the terminal voltage sags by I·R0 discharging and rises by it charging, which is simultaneously the per-direction efficiency (eq 18.5), the charge-acceptance taper (eq 18.6), and the Joule heat driving self-heating (eq 18.10). R0 is never a guess: it is derived from the sheet's own printed round-trip efficiency at its reference rate (eq 18.4). Original figure; standard equivalent-circuit (Rint) topology.

Worked example — the internal resistance a 96% sheet implies

A 10 kWh, 51.2 V lithium iron phosphate pack prints 96% round-trip efficiency. Coulombic efficiency for lithium is 0.998, so the voltage share is rv = 0.96 / 0.998 = 0.9619; the resistive ratio is (1 − 0.9619)/(1 + 0.9619) = 0.0194. The reference rate is 0.5C — 10,000 Wh / 51.2 V × 0.5 = 97.7 A — so R0 = 0.0194 × 51.2 / 97.7 = 10.2 mΩ (eq 18.4). Discharge at that reference rate and the terminal sags 97.7 A × 10.2 mΩ ≈ 1.0 V — about 2% of pack voltage each way, which round-trips to exactly the printed 96%: the sheet's own number, reproduced rather than assumed.

state of charge (SOC)0%100%rest voltage, as a position in each chemistry's own windowLFP — flat plateauNMC — steady slopelead-acid — near-linear
Fig. 18.2 — Chemistry sets the rest-voltage SHAPE — each curve normalized to its own voltage window so the forms compare directly. Lithium iron phosphate holds a famously flat mid-range plateau (its voltage barely moves from 20% to 90% state of charge), lithium NMC slopes steadily, and lead-acid is nearly a straight line. These shapes drive everything downstream: a flat plateau keeps charge-acceptance headroom to the very top (Fig. 18.3's short LFP taper) and makes voltage a poor fuel gauge; a slope does the opposite. Drawn from the model's own per-cell anchor tables (Springer CJME 2018 / ACCURE class data).
state of charge (SOC)0%100%accepted charge, as a share of the charger capconstant current — the cap bindsNMC taperlead-acid absorption taperLFP holds to full, then the window cuts
Fig. 18.3 — Constant-current / constant-voltage charging emerges from the circuit rather than being scripted (eq 18.6). While the voltage headroom above the rest voltage exceeds the charger's current cap, the cap binds — the flat constant-current region. Where the rising rest voltage eats that headroom, acceptance tapers — the constant-voltage region. Lithium iron phosphate's flat plateau (Fig. 18.2) keeps headroom essentially to the top (the famously short LFP taper, ending at the battery-management window's hard cut at full); NMC tapers just below full; lead-acid's rising voltage and absorption polarization start its long taper around 60% — the classic slow lead-acid finish. Drawn from the model's own acceptance function at 25 °C, each chemistry at its named charge convention (0.5C lithium / 0.2C lead-acid bulk).

The same circuit gives the rest: deliverable power is sag-limited (the voltage headroom above the cutoff, over R₀), lead-acid’s famously slow absorption falls out of its own rising charge resistance, lithium refuses to charge below 0 °C (a plating-protection reality enforced by the battery management system, BMS), and the BMS window terminates charge exactly at full. Losses, unabsorbed solar and unserved demand are all echoed per hour — conservation is exact by construction.

SymbolMeaningUnitsValue / source
ηcCoulombic efficiency—0.998 lithium / 0.90 lead-acid (Battery University BU-808c classes; datasheet RTE always wins the split)
VchgCharge-stage voltage per cellV3.65 LFP / 4.20 NMC / 2.40 lead-acid — universal charger/BMS conventions, scaled to the pack
IrefReference rateA0.5C lithium; the C10 rate for lead-acid (the conditions printed figures assume, named in the response)
nPeuPeukert exponent—1.25 lead-acid (published 1.2–1.5) / ~1.02 lithium; clipped 0.5–1.05× rated capacity (honest cap)

Aging is counted, not assumed. Cycle wear comes from a rainflow decomposition of the year’s actual state-of-charge swings under Miner’s rule, with every stress shape anchored at the sheet’s printed cycle-life point — so all factors equal 1 at the rating condition and a rating-condition cycle costs exactly 1/Ncyc:

High-rate hours therefore cost MORE than their energy share (Wang et al. 2011: ≈ +16% fade per +1C at 25 °C; the NMC class about twice as rate-sensitive) — the response echoes that rate-stress multiplier explicitly, which is the surface a future grid-coupled installation reads. Calendar fade follows square-root-of-time kinetics (Naumann et al. 2018) stressed by mean state of charge and temperature. The reported outlook is a 25-year TRAJECTORY, not a single number: each year’s fade is computed on the already-faded pack (year 1’s usage recurring in absolute terms), so degradation compounds the way the knee literature’s snowball class describes (Attia et al. 2022) — and the years-to-80% figure comes from that trajectory, echoed beside its no-feedback comparison value. Either figure is presented no further than a 40-year horizon: past that the square-root extrapolation has left its validated domain (real lithium shelf life sits in the 15–25-year class), so the display says “more than 40 years” instead of printing arithmetic as a forecast.

The pack heats itself, and aging feels it. The hourly Joule loss the energy ledger already books drives a lumped thermal node (solved in closed form each step — unconditionally stable, exact fixed point):

(18.4) Internal resistance from the printed round trip (the datasheet-anchoring move)
rv=RTEηc,R0=1−rv1+rv⋅VnomIref{\htmlData{sym=r_v}{r_v}} = \dfrac{{\htmlData{sym=RTE}{\text{RTE}}}}{{\htmlData{sym=eta_c}{\eta_c}}}, \qquad {\htmlData{sym=R_0}{R_0}} = \dfrac{1 - {\htmlData{sym=r_v}{r_v}}}{1 + {\htmlData{sym=r_v}{r_v}}}\cdot\dfrac{{\htmlData{sym=V_nomb}{V_{\text{nom}}}}}{{\htmlData{sym=I_refb}{I_{\text{ref}}}}}
(18.5) Per-direction step efficiencies (reference-rate product closes on the printed RTE — identity anchor)
ηchg=OCV ηcOCV+I R0,ηdis=OCV−I R0OCV{\htmlData{sym=eta_chg}{\eta_{\text{chg}}}} = \dfrac{{\htmlData{sym=OCV}{\text{OCV}}}\,{\htmlData{sym=eta_c}{\eta_c}}}{{\htmlData{sym=OCV}{\text{OCV}}} + {\htmlData{sym=I_bat}{I}}\,{\htmlData{sym=R_0}{R_0}}}, \qquad {\htmlData{sym=eta_dis}{\eta_{\text{dis}}}} = \dfrac{{\htmlData{sym=OCV}{\text{OCV}}} - {\htmlData{sym=I_bat}{I}}\,{\htmlData{sym=R_0}{R_0}}}{{\htmlData{sym=OCV}{\text{OCV}}}}
(18.6) Charge acceptance (CC–CV emerges: headroom = current limit binds; vanishing headroom = the taper)
Iacc=Vchg−OCV(SOC)R0(Tbat){\htmlData{sym=I_acc}{I_{\text{acc}}}} = \dfrac{{\htmlData{sym=V_chg}{V_{\text{chg}}}} - {\htmlData{sym=OCV}{\text{OCV}}}({\htmlData{sym=SOC}{\text{SOC}}})}{{\htmlData{sym=R_0}{R_0}}({\htmlData{sym=T_bat}{T_{\text{bat}}}})}
(18.7) Rate-effective capacity (Peukert 1897 — first-order lead-acid, near-flat lithium)
Qeff=Qrated (Iref/I) nPeu−1{\htmlData{sym=Q_effb}{Q_{\text{eff}}}} = {\htmlData{sym=Q_rated}{Q_{\text{rated}}}}\,\big({\htmlData{sym=I_refb}{I_{\text{ref}}}}/{\htmlData{sym=I_bat}{I}}\big)^{\,{\htmlData{sym=n_Peu}{n_{\text{Peu}}}} - 1}
(18.8) Cycle-life consumption (rainflow + Miner; identity-anchored at the printed rating point)
Fcyc=∑ici (di/DoDr)kpNcyc  frate ftemp{\htmlData{sym=F_cyc}{F_{\text{cyc}}}} = \dfrac{\sum_{\htmlData{sym=i_cyc}{i}} {\htmlData{sym=c_i}{c_i}}\,\big({\htmlData{sym=d_ib}{d_i}}/{\htmlData{sym=DoD_r}{\text{DoD}_r}}\big)^{{\htmlData{sym=k_pb}{k_p}}}}{{\htmlData{sym=N_cyc}{N_{\text{cyc}}}}}\; {\htmlData{sym=f_rate}{f_{\text{rate}}}}\, {\htmlData{sym=f_tempb}{f_{\text{temp}}}}
(18.9) Stress factors + calendar fade (throughput-weighted means over the year)
frate=⟨e kC Crate⟩e kC Cref,ftemp=⟨e−Ea,cycRgas(1Tbat−1298)⟩,fadecal=kcaltyr{\htmlData{sym=f_rate}{f_{\text{rate}}}} = \dfrac{\langle {\htmlData{sym=e_euler}{e}}^{\,{\htmlData{sym=k_C}{k_C}}\,{\htmlData{sym=C_rate}{C_{\text{rate}}}}}\rangle}{{\htmlData{sym=e_euler}{e}}^{\,{\htmlData{sym=k_C}{k_C}}\,{\htmlData{sym=C_refr}{C_{\text{ref}}}}}}, \quad {\htmlData{sym=f_tempb}{f_{\text{temp}}}} = \Big\langle {\htmlData{sym=e_euler}{e}}^{-\frac{{\htmlData{sym=E_a_cyc}{E_{a,\text{cyc}}}}}{{\htmlData{sym=R_gas}{R_{\text{gas}}}}}\left(\frac{1}{{\htmlData{sym=T_bat}{T_{\text{bat}}}}} - \frac{1}{298}\right)}\Big\rangle, \quad \text{fade}_{\text{cal}} = {\htmlData{sym=k_cal}{k_{\text{cal}}}}\sqrt{{\htmlData{sym=t_yr}{t_{\text{yr}}}}}
(18.10) Self-heating node (Bernardi irreversible term; entropic term cancels over any full cycle, named)
Cth dTbatdt=Q˙loss−hA (Tbat−Tenv){\htmlData{sym=C_th}{C_{\text{th}}}}\,\dfrac{{\htmlData{sym=d_diff}{\mathrm{d}}}{\htmlData{sym=T_bat}{T_{\text{bat}}}}}{{\htmlData{sym=d_diff}{\mathrm{d}}}{\htmlData{sym=t_time}{t}}} = {\htmlData{sym=Qdot_loss}{\dot{Q}_{\text{loss}}}} - {\htmlData{sym=hA}{hA}}\,\big({\htmlData{sym=T_bat}{T_{\text{bat}}}} - {\htmlData{sym=T_env}{T_{\text{env}}}}\big)
(18.11) Multi-year trajectory — fade feeds back (25-year horizon)
SOHk+1=SOHk−[Fcyc(SOHk)+kcal(k+1−k)]{\htmlData{sym=SOH}{\text{SOH}}}_{{\htmlData{sym=k_yr}{k}}+1} = {\htmlData{sym=SOH}{\text{SOH}}}_{\htmlData{sym=k_yr}{k}} - \big[ {\htmlData{sym=F_cyc}{F_{\text{cyc}}}}({\htmlData{sym=SOH}{\text{SOH}}}_{\htmlData{sym=k_yr}{k}}) + {\htmlData{sym=k_cal}{k_{\text{cal}}}}\big(\sqrt{{\htmlData{sym=k_yr}{k}}{+}1} - \sqrt{{\htmlData{sym=k_yr}{k}}}\big) \big]

Resistance, charge acceptance, the 0 °C lithium block and both aging stress factors all evaluate at Tbat — never at raw ambient. An optional outdoor enclosure closes the loop outward: each wall and the roof receives its own computed sunlight through the SAME transposition the panels use (§5), driving a single-zone envelope model whose interior becomes the pack’s environment, with passive phase-change buffering sized honestly against its overnight regeneration. Opt-in active options complete the picture (eqs 19.14–19.15): a night fan whose flow IS conductance (0.34 W/K per m³/h — air's own heat capacity), an air-source heat pump at the conservative COP 3 class holding the setpoints with any unmet hours reported, the hybrid running the fan first and the pump only when needed, and cold-climate capture of the panels' own waste heat (unglazed photovoltaic-thermal class — the measured pilots also gained ≈4% annual electricity from the cooling side, an effect noted but not yet fed back into the module chain). And for the fully buried option (eq 18.16), the ground itself becomes the heat sink: soil temperature at depth is a damped, phase-lagged copy of the yearly air wave (Kusuda–Achenbach — the EnergyPlus-standard model, computed entirely from this site's own temperatures), so at ~3 m a buried enclosure lives near the annual mean year-round — dissolving both the hot-climate saturation problem and, wherever the annual mean sits above freezing, the lithium cold-charge block. The excavation, reinforcement, drainage and ventilation that make burial safe are real costs this model deliberately does not include:

Gface — each face's OWN transposed sunlight (§5)ΔRlw — roof radiates to the cold skyenvelope U · A — bare steel vs foam insulation classesTsa per facebatteryQ̇loss — its own Joule heatphase-change slabmelts at its set pointinterior Tenvnight rejectionre-freezes the slabfloor adiabatic (named)
Fig. 18.5 — The enclosure as a single thermal zone (eq 18.12). Outside, every face experiences its own sol-air temperature — ambient plus its absorbed sunlight over the surface film coefficient, minus the roof's longwave loss to the cold sky — with each face's sunlight computed by the SAME transposition the panels use (§5): a dark west wall and the roof see very different hours. Heat conducts through the envelope in proportion to its U-value (bare steel ≈ 5.6 W/m²K from the film coefficients alone; foam insulation cuts it ~20×). Inside, the battery adds its own Joule heat, and an optional phase-change slab absorbs excess by melting at its chosen temperature — re-freezing overnight only if night air can reach below its melt point, which is why the sizing echo carries a regeneration margin. Original figure; sol-air construction per ASHRAE.

Worked example — how much phase-change material one hot day needs

Suppose the worst summer day pushes 30 kWh of heat into the enclosure above the 25 °C setpoint (envelope conduction under sol-air forcing plus the battery's own losses). At a 200 kJ/kg latent-heat class, 30 kWh × 3,600 kJ/kWh ÷ 200 kJ/kg = 540 kg of phase-change material to hold that day passively (eq 18.13). Whether 540 kg keeps working is the regeneration question: the night must reject those same 30 kWh back through the envelope below the melt point, or tomorrow starts with a spent buffer — the echoed overnight regeneration margin says which, and insulation cuts BOTH numbers (~20× from bare steel to 100 mm foam) before any material is bought.

(18.12) Sol-air forcing per enclosure face (ASHRAE construction)
Tsa=Tamb+αsol Gfaceho−ΔRlw{\htmlData{sym=T_sa}{T_{\text{sa}}}} = {\htmlData{sym=T_amb}{T_{\text{amb}}}} + \dfrac{{\htmlData{sym=alpha_sol}{\alpha_{\text{sol}}}}\,{\htmlData{sym=G_face}{G_{\text{face}}}}}{{\htmlData{sym=h_o}{h_o}}} - {\htmlData{sym=dR_lw}{\Delta R_{\text{lw}}}}
(18.13) Worst-day phase-change sizing (honest only beside its overnight regeneration margin)
mpcm=Eexchfus{\htmlData{sym=m_pcm}{m_{\text{pcm}}}} = \dfrac{{\htmlData{sym=E_exc}{E_{\text{exc}}}}}{{\htmlData{sym=h_fus}{h_{\text{fus}}}}}
(18.14) Forced-air night regeneration (optional) — flow IS conductance
UAvent=0.34 V˙air{\htmlData{sym=UA_vent}{UA_{\text{vent}}}} = 0.34\, {\htmlData{sym=Vdot_air}{\dot{V}_{\text{air}}}}
(18.15) Active heat pump (optional) — electricity per unit heat moved
Eel=Qmoved/COP{\htmlData{sym=E_el}{E_{\text{el}}}} = {\htmlData{sym=Q_moved}{Q_{\text{moved}}}} / {\htmlData{sym=COP}{\text{COP}}}
(18.16) Geothermal burial (optional) — Kusuda–Achenbach undisturbed ground temperature
Tgnd(zb,t)=Tˉair−As e−zb/dg cos⁡ ⁣(2π (t−tcold)8760−zbdg){\htmlData{sym=T_gnd}{T_{\text{gnd}}}}({\htmlData{sym=z_b}{z_b}}, {\htmlData{sym=t_time}{t}}) = {\htmlData{sym=Tbar_air}{\bar{T}_{\text{air}}}} - {\htmlData{sym=A_s}{A_s}}\,{\htmlData{sym=e_euler}{e}}^{-{\htmlData{sym=z_b}{z_b}}/{\htmlData{sym=d_g}{d_g}}}\,\cos\!\Big(\tfrac{2{\htmlData{sym=pi_const}{\pi}}\,({\htmlData{sym=t_time}{t}} - {\htmlData{sym=t_cold}{t_{\text{cold}}}})}{8760} - \tfrac{{\htmlData{sym=z_b}{z_b}}}{{\htmlData{sym=d_g}{d_g}}}\Big)
SymbolMeaningUnitsValue / source
hoExterior film coefficientW/(m²·K)17 — the ASHRAE conservative standard; bare-steel envelope U ≈ 5.6 from the film coefficients alone, foam classes add their thickness/conductivity
ΔRlwLongwave sky correction°C≈3.9 for sky-facing roofs, ~0 for walls (ASHRAE — the low effective sky temperature)
αsolSkin absorptivity—0.45 light paint / 0.90 dark (cited conservative pair)
hfusPhase-change latent heatkJ/kg200 class (paraffin/salt-hydrate range 150–250); the melt point must sit inside the diurnal band or the buffer never regenerates — the margin is echoed
CthPack thermal capacitanceJ/Kspecific heat × mass — lithium cells ~1,000–1,100 J/(kg·K) (Battery Design measured compilations); datasheet weight wins the mass, chemistry-class kg/kWh fills otherwise

What the response reports: strings and conversion blocks per array, string_windows for the voltage-window verdicts, battery_autonomy (monthly recharge capability, storable share, runtime at constant draws, the aging outlook), dispatch with a load profile (self-sufficiency, the unmet-energy series and backup-generator sizing — including the standard-catalog generator fit: the smallest commercial kVA unit covering the peak deficit and, when different, the smaller unit covering 99% of hours at healthier engine load, with fuel from the published part-load consumption class and sustained sub-30%-load running flagged as a wet-stacking risk; the daily load shape you enter is local clock time — the model aligns it to the weather data’s UTC hours from the site’s longitude, the same convention the sun-position calculation uses), and enclosure when one is configured — each absent entirely when its input is not supplied, and each carrying its full assumption list verbatim. The battery core has been externally refereed against NREL’s SAM battery model (PySAM): voltage shapes, the CC–CV taper, the thermal form and the cycle-damage shape all reproduce; remaining level differences trace to the anchor choice — datasheet figures here, NREL cell defaults there — which is this tool’s deliberate doctrine.

Reference

19. Global Heatmap Overlays (Map Layers)▾

The map's selectable layers are pre-computed climatology rasters, not live simulation output: NASA POWER MERRA-2 1981–2020 climatology at 1° resolution, served as a single static file. They are for site comparison — the per-site simulation uses the (finer, more recent) sources in the Data Sources table below.

LayerNASA POWER variableMetricNote
Annual Yield (TMY) / GHIALLSKY_SFC_SW_DWNkWh/m²/yrdaily mean × 365
TemperatureT2M°Cannual mean
RainfallPRECTOTCORRmm/yrdaily mean × 365; √ colour scale
SnowfallPRECSNOmm/yr water equiv.γ = 0.35 power colour scale
WindWS2Mm/sannual mean at 2 m — ≈ module height (the simulation itself uses 10 m wind, §8)
Elevationtile geometrymMERRA-2 DEM; √ colour scale
Yield variabilityprecomputed CV grid (§16)CV %cv at ×1 soiling, aggregated 0.5° → 1°

Colour ramps use power/gamma normalisation where linear scaling would crush the interesting range (e.g. snowfall γ = 0.35 keeps ephemeral-snow regions visible without flattening Scandinavia; rainfall and elevation use γ = 0.5). The wind layer deliberately shows 2 m wind — roughly module height, honest for site browsing — while the Faiman model uses 10 m wind for consistency with PVGIS.

Concept — why the colour scales are bent

Most climate variables are lopsided: nearly all the world's snowfall range belongs to a few extreme regions, while the differences that matter for a solar decision — none, a little, some — are crowded into the bottom few percent. A linear colour ramp would spend almost all of its colours on Greenland and render an ephemeral-snow site the same colour as a desert. The power scale bends the mapping (Fig. 19.1): values are raised to γ < 1 before colouring, stretching the low end apart and compressing the extreme top — the cartographic cousin of plotting on a log axis.

25%50%75%100%0%25%50%75%100%data value (fraction of the layer range)position on the colour ramplinearγ = 0.35~50 mm/yr: 1% of the range → 20% up the ramp
Fig. 19.1 — Gamma normalisation, drawn with the live snowfall exponent (γ = 0.35). On a linear scale (dashed), a region with ~50 mm/yr of snow — 1% of the 5,000 mm/yr range, but enough to matter for yield — would get 1% of the colour ramp: invisible. The power curve stretches the bottom of the scale, lifting that same 1% to ≈20% of the ramp, while the far high end compresses — Scandinavia stays distinguishable, and ephemeral-snow regions stop disappearing. Original figure.
Statistical Methods — Summary▾

Every statistical process used anywhere in the model, what it does, and where it applies. Details live in the referenced sections; this table is the single checklist.

ProcessUsed inWhat it does / why it is valid
Typical Meteorological Year construction (Finkelstein–Schafer distribution matching)§2, §16Selects, per calendar month, the historical month whose climate-variable distributions best match the 16-year record — a distribution-representative composite year (the P50 anchor), not an average.
Climatological normals (multi-year monthly means)§8 wind, §3 fallback temp, §10 rain + AOD, §11 snowLong-window monthly averages smooth out single-year anomalies; window lengths are chosen per variable to span multiple ENSO cycles (§15).
Monthly → hourly downscaling§8, §3Monthly wind is broadcast to every hour of its month; fallback-path temperature is circularly interpolated between mid-month anchors with a ±3 °C diurnal cosine (peak 14:00).
Triangular / flat-top trapezoid cycle averaging§10Closed-form mean of a linear accumulate-then-reset sawtooth (peak/2), switching to a trapezoid mean once the 30% saturation cap truncates the peak.
Sample coefficient of variation, CV = σ/μ (ddof = 1)§16Relative spread of 39 simulated annual yields per grid cell; the sample (n−1) standard deviation corrects small-sample bias.
Normal-quantile downside bounds (z = 1.28, 0.674)§16P90 = P50(1 − 1.28·CV), P75 = P50(1 − 0.674·CV) — assumes annual yields are approximately normally distributed about the median.
Joint historical simulation§16All climate variables vary together per year, preserving real correlations (rain ↔ cloud ↔ temperature) that independent worst-case stacking would ignore.
Root-sum-square (RSS) combination§16 snow patchIndependent variances add: cv_total = √(cv_climate² + cv_snow²). Valid because snow (Arctic/North Atlantic Oscillation) and climate (ENSO/PDO) variability have largely orthogonal drivers — the two records need not share year windows.
Bias cancellation in ratios§16The grid simulation runs at monthly resolution, which biases absolute yield — but the bias is constant year-to-year, scaling σ and μ equally, so it cancels in CV. The absolute scale comes from PVGIS instead.
Proxy variable (GHI for plane-of-array)§16The grid uses GHI variability as a stand-in for tracker plane-of-array variability — the (DNI+DHI)/GHI ratio is roughly constant year-to-year at a site, leaving CV nearly unchanged (marginally conservative).
Nearest-neighbour lookup§16Site coordinates snap to the nearest 0.5° cell centre and the soiling multiplier to the nearest precomputed variant — no spatial interpolation is applied.

Sampling uncertainty: a 39-year record (≈1.3 PDO cycles) estimates CV to roughly ±12% relative uncertainty; the 19-year snow record is coarser (~±15–20%). These are uncertainties on the spread estimate itself, second-order relative to the spread it measures. The per-cell "representative low-yield year" shown on the P90 card is the historical year whose simulated yield lies closest to the P90 target.

What Is Not Yet Modelled▾
  • System-layer residuals — the composed hourly array energy, string wiring, inverter conversion and voltage-window checks are live (§18). Still deferred: explicit per-string panel lists (strings group in fill order), a state-of-charge-coupled PWM bus voltage (the charge-band midpoint is used, stated), part-load inverter efficiency at the combined battery-plus-load flow on the load-serving side (a flat weighted figure applies there), and weekly/seasonal load-shape variation (one day shape tiles the year, stated in the dispatch assumptions).
  • P90 grid and module parameters — the interannual CV grid (§16) was precomputed with the generic module and the earlier uniform-sky transposition. Efficiency scales out of a CV exactly (it multiplies every year's yield equally); the module thermal parameters and the sky-model change shift every year's yield near-equally, so both move the CV only at second order, well inside the CV's own sampling uncertainty — the grid is applied unchanged.
  • Bifacial second-order terms — the §17 named list: rear soiling, single-module rear mismatch (below 0.5%), and the 2-D cross-section geometry (quantified and justified in §17 — conservative, and the same class the bankable tools use). (The snow-albedo lift reaches every surface, the tilt sweep optimises the bifacial yield, and spectral albedo is now modelled statically via the §2 per-surface factor — what remains spectral is hourly variation and unmeasured surface spectra, e.g. asphalt.)
  • Bypass-diode behaviour under shade — live with the §4 near-field path (API near_objects): c-Si modules with the electrical block run the exact substring solve with the diode clamp, ideal-MPPT. Still deferred: real-MPPT multi-peak P–V behaviour and Bishop hot-spot analysis. The hourly voltage/current series the curve model computes are also not currently exposed as a chart.
  • Shading & obstructions — both regimes are live end-to-end (§4: far-field horizon gate + sky-view factors; near-field strip shade through the §9.3 diode tiers), with the site editor supplying the inputs (envelope, wall lines, horizon arcs and terrain fetch, phone scan, corner heights). Remaining: obstruction reflectance and penumbra, the bifacial rear sky-strip correction, the tilt-sweep / αr-table near-field pass, and two quantified array diffuse-side terms — inter-row ground-reflected dimming is not modelled (≈0.9% of yield at a very dense 0.57 ground-cover ratio, near zero at normally-spaced rows; cross-checked against pvlib's infinite-sheds model, 2026) and per-panel sky masking is evaluated at the module centre (reads ~half the collector-integrated value at that same stress density). The array beam-shading term itself matches the infinite-sheds beam component to 0.05 percentage points.
  • Battery residuals — charge/discharge, aging with the 25-year fade-feedback trajectory, self-heating, the active heat-exchanger options (night fan, heat pump, hybrid, panel-heat capture) and geothermal burial are all live (§18). Still deferred: diesel-generator and grid flows as modelled ENERGY SOURCES in the balance itself (the standard-catalog generator fit — unit selection, runtime, fuel and wet-stacking risk — is computed from the unmet-energy series, but generator output does not yet feed back into the served-energy accounting, and grid import/export is not modelled), per-product measured temperature/aging curves (chemistry-class defaults apply when a sheet does not print them, provenance echoed), served-energy feedback in the fade trajectory (the faded pack also delivers less — stress-side feedback only today), the panel-cooling electrical lift from heat capture (echoed as an estimate, not fed into the module chain), knee-onset prediction (plating/saturation pathways have no datasheet basis), and cost-optimal control of the hybrid thermal options (the physics books each option's energy; choosing between them by price is planned financial-analysis work).
  • Financial model — no LCOE, payback, or revenue calculation (planned).
  • Log-law wind height correction — simulation uses 10 m wind; module height is ~1.5 m. A roughness-length correction is planned.
  • Cross-array shading — inter-row and inter-pedestal shading are fully modelled WITHIN each array (§4); what separate arrays of a split system cast on each other is not (they are treated as disjoint faces — echoed in every multi-array response).
  • Soiling cementation — wet dust hardening (gypsum, halite) reduces cleaning effectiveness but is not yet modelled. Most relevant at high multiplier settings (mine site, industrial).
  • Spectral effects — efficiency assumed spectrally flat for sky and beam light (reasonable for broadband irradiance); ground-reflected light carries the §2 per-surface × per-technology spectral factor.
  • P90 data/model uncertainty terms — bankable assessments add irradiance-dataset and model uncertainty (~3–5% each) to the interannual spread via RSS; only the interannual + snow terms are currently stacked.
  • Horizon brightening & rear circumsolar — sky diffuse transposes by Hay–Davies (§5): circumsolar is modelled, the Perez-class horizon-brightening band is not (a modelled brightness band would double-count sky this tool's measured horizon profiles treat explicitly — deliberate). The bifacial rear sees the isotropic remainder only; rear-side circumsolar during sun-behind-panel hours is not added (conservative). Residual vs Perez-based tools (PVsyst, SAM): ~0–3% either direction (verified at six climates vs pvlib and NREL SAM, 2026).
Data Sources — APIs, Variables, Units & Conversions▾

Everything the simulation fetches, exactly as fetched: the API variable, its native unit, and any conversion applied before use. All APIs are free and unauthenticated.

DataSource (dataset)Variable · native unitsConversion appliedWindowResolution
Irradiance (hourly)PVGIS TMY (SARAH-3/ERA5)G(h), Gb(n), Gd(h) · W/m²none — used directly2005–2020 (16-yr TMY)~5 km
Ambient temperature (hourly)PVGIS TMYT2m · °Clapse-rate corrected to site elevation (§3)2005–2020 (16-yr TMY)~5 km
Temperature (fallback path)Open-Meteo ERA5temperature_2m_mean · °C/daydaily → 12 monthly means → hourly (§3)2018–2022~11 km
RainfallNASA POWER MERRA-2 climatologyPRECTOTCORR · mm/day× days-in-month (Feb 28.25) → mm/month2001–2020~50 km
Wind speedNASA POWER MERRA-2 climatologyWS10M · m/s at 10 mclipped to ≥ 0.5 m/s; monthly → hourly broadcast2001–2020~50 km
Surface albedo (advisory + seasonal mode)NASA POWER MERRA-2 climatologyALLSKY_SRF_ALB · dimensionlessvalues outside (0, 1] (fill −999) disable it; seasonal mode: monthly shape × site level over snow-free months, clipped 0.05–0.9 (§2)2001–2020~50 km
SnowfallOpen-Meteo ERA5snowfall_sum · cm water-equiv./daydaily → mean daily per month; × days-in-month in the model2004–2022~11 km
Aerosol optical depthNASA POWER SYN1DEGAOD_55 · dimensionless−999 sentinels + month-13 annual rows filtered; monthly means across years; missing months → annual mean (flagged)2018–2022~55 km
Site elevationOpen-Topo-Data (ASTER GDEM)metres (integer)none; Open-Elevation SRTM fallback (~90 m; silently 0 above ~60° latitude)static~30 m
Map overlaysNASA POWER MERRA-2 climatologyALLSKY_SFC_SW_DWN, T2M, PRECTOTCORR, PRECSNO, WS2M (2 m)annual means; totals × 3651981–20201° (~111 km)
P90 CV gridNASA POWER monthly (offline, 39-yr joint sim)GHI · T2M · rain · wind · AOD (§16)CV per cell × 4 soiling variants1984–20220.5° (~55 km)
Snow CV (P90 RSS patch)Open-Meteo ERA5 (offline)snowfall_sum per-year monthly totalsannual snow-loss fractions → cv_snow2004–20220.5° cells

Silent unit mismatches are treated as a first-class failure mode here: the wind km/h incident (§8, §15) cost a 3.7% yield error before it was caught, so every row above states the native unit explicitly.

Glossary — Symbols & Parameters▾

Every variable, parameter and constant that appears in the equations above and in the model-map subsystems, taken from the source. Most Greek symbols show up only in the subsystem detail views (geometry, incidence, thermal) rather than the headline equations, which is why they may look unfamiliar.

Greek symbols

SymbolMeaningUnitsUsed in
ηcCharge that comes back out per charge put in (amp-hour basis): ~0.998 lithium, ~0.90 lead-acid (Battery University BU-808c classes) — the non-resistive share of round-trip loss—§18
ηchgEnergy stored per unit absorbed at one operating point (eq 18.5) — falls as current rises because the terminal sits I·R₀ above the rest voltage; the charge–discharge product at the reference rate closes exactly on the printed RTE (identity anchor)—§18
ηdisEnergy delivered per unit drawn at one operating point (eq 18.5) — the terminal sits I·R₀ below the rest voltage under load—§18
αsolFraction of incident sunlight the enclosure skin absorbs: 0.45 light paint / 0.90 dark (cited conservative pair). Subscripted — α_p is the power temperature coefficient and α_r the angular-loss coefficient—§18
ΔRlwSol-air reduction for radiation to the cold sky: ~3.9 °C for sky-facing roofs, ~0 for vertical walls (ASHRAE) — why a roof runs cooler than its sunlight alone suggests, and why it keeps cooling at night°C§18
ηinvDC→AC conversion efficiency at the hour's load ratio (eq 18.3, the PVWatts part-load form, NREL/TP-6A20-62641) — low mornings pay the fixed electronics overhead, the flat top runs near nominal. Subscripted: base η is module efficiency—§18
ηnomThe part-load curve's scale, anchored so the CEC-weighted average of the modelled curve equals the sheet's printed WEIGHTED efficiency (an energy statement, weights 4/5/12/21/53/5% at 10–100% load) — or the printed curve itself when supplied—§18
ζHourly DC input as a fraction of the inverter's rated DC power — the part-load curve's argument (eq 18.3)—§18
αr,effα_r interpolated piecewise-linearly at the current mean soiling level (§10.3 breakpoints)—§10
ηrel,iRelative efficiency at the i-th irradiance anchor S_i — the datasheet point when supplied, else the technology-typical fallback value; eq 9.2 interpolates between consecutive anchors (η_rel,i, η_rel,i+1)—§9
αIscShort-circuit-current temperature coefficient — datasheet value wins, per-technology median pre-fills (c-Si +0.05 %/°C); resolved absolute [A/K] in the translation (eq 9.5)%/°C → A/K§9.3
βVocOpen-circuit-voltage temperature coefficient — the fifth extraction constraint (central ±1 K difference of the model's own Voc(T)). Datasheet value wins, per-technology median pre-fills (c-Si −0.30 %/°C); guarded −0.6 to −0.05 %/°C%/°C → V/K§9.3
λstdStandard meridian of the site's time zone = round(λ/15)×15°§1
ρsnowGround albedo with snow cover (0.65)—§11
ρclearGround albedo without snow cover (0.20) in the albedo-bonus formula—§11
φsSolar azimuth — bearing from north, negative east / positive west. φ is reserved for azimuths; latitude is L (see notation note in §1)°§1, §6
φpPanel azimuth — 0° (due north) in the southern hemisphere, 180° (due south) in the northern°§6
αrAngular-loss coefficient (Martín–Ruiz); clean anchor 0.17 (0.14 for AR-coated glass, or a measured module value 0.12–0.19) → 0.27 heavily soiled—§5, §6, §10
βSolar elevation angle°§1, §5, §6, §4
γModule power temperature coefficient (default −0.004; the datasheet γ_Pmax in %/°C ÷ 100 when module parameters are set)/°C§9
δSolar declination°§1
ηrel,200Datasheet relative efficiency at the IEC 61853-1 low-irradiance test point (200 W/m², 25 °C) — e.g. 0.965 for a −3.5% low-light spec—§9
ηrelIrradiance-dependent efficiency multiplier — piecewise log-linear through whichever datasheet points (200–800 W/m²) are supplied plus the definitional STC point; with only the 200 W/m² value it reduces to a one-parameter log fit; with no datasheet points, the measured technology-typical curve for the chosen cell type applies (never flat)—§9
ηModule efficiency at STC (default 0.20; derived from datasheet rated power ÷ (area × 1000 W/m²) when module parameters are set — derived value guarded to 5–30%)—§9, §14
θAngle of incidence (fixed-tilt beam; in §4's backtracking cost it is the tracker's beam incidence after rotating ψ off the sun)°§6, §4
ψAzimuth rotation off sun-normal tracking — the smallest |ψ| (within ±60°) that makes every pair overlap vanish; the incidence cost cos θ = cos²β·cos ψ + sin²β is even in ψ, so minimal rotation is provably minimal cost. Infeasible hours keep the overlap-minimising ψ with the residual echoed°§4
ΔshIn-plane displacement of a neighbour aperture's shadow on the receiver plane — parallel planes project a rectangle to a CONGRUENT translated rectangle, so shading reduces to 1-D interval overlaps (Lorenzo, Narvarte & Muñoz 2011)m§4
λSite longitude, east positive°§1
ρGround albedo — fraction of GHI reflected onto the panel; site ground-surface input (§2 table), default 0.20—§2, §11
ΣPanel tilt / surface slope angle — also the summation operator in §7 and §14 (with a running index, Σt)°§5–§7, §10, §11
σStandard deviation of annual yields (in CV = σ / mean)kWh/m²§16
αAltitude coordinate of a sky-dome element in the §4 view-factor integrals — 0° at the horizon, 90° at the zenith. Not the soiling angular-loss coefficient α_r, which always wears its subscript°§4
φAzimuth coordinate of a sky-dome element — paired with the altitude α to sweep the whole dome in the view-factor integrals°§4

Variables & functions

SymbolMeaningUnitsUsed in
TgndSoil temperature at the burial depth per Kusuda–Achenbach (the EnergyPlus-standard model, eq 18.16) — a damped, phase-lagged copy of the yearly air-temperature wave, computed from this site’s own series. At ~3 m it is essentially the annual mean; a buried enclosure exchanges with THIS instead of sol-air. Honest limit: a site whose annual mean is below zero has ground below zero — burial does not conjure heat°C§18
zbDepth of soil cover over the buried enclosure — default 3 m (past the annual damping depth; a physics default, overridable). Deeper is flatter and more phase-lagged, with rapidly diminishing returns beyond ~2× the damping depthm§18
dgThe e-folding depth of the annual temperature wave in soil = √(2α/ω) — ≈2.6 m at the moist-soil diffusivity class (0.7×10⁻⁶ m²/s). Each damping depth of cover shrinks the annual swing by e (≈63%)m§18
AsHalf the swing between the warmest and coldest monthly-mean air temperatures at the site — the amplitude the ground wave damps (eq 18.16); taken from the site’s own series°C§18
tcoldHour of the year the surface wave bottoms out — sets the ground wave’s phase; the lag z_b/d_g means the ground is coldest WEEKS after the air (why cellars feel cool in early summer)h§18
T̄airThe site’s annual mean — what deep ground converges to (eq 18.16), and therefore what a well-buried enclosure lives at year-round°C§18
COPHeat moved per unit of electricity a heat pump consumes — air-source class 3.0 here (conservative; published 3–4), applied both heating and cooling. The enclosure books E = Q/COP whenever the active option runs (eq 18.15)—§18
V̇airFan-driven air flow through the enclosure — every m³/h of flow carries 0.34 W/K of conductance to ambient (air’s ρ·c_p), which is what re-freezes the phase-change slab overnight (eq 18.14); fan electricity is booked at the ~0.4 W per m³/h specific-fan-power classm³/h§18
UAventThe extra enclosure-to-ambient conductance the night fan adds — 0.34 × flow (eq 18.14). A 1,000 m³/h fan adds ≈340 W/K, dwarfing even a bare-steel envelope’s ~330 W/K — which is why night regeneration worksW/K§18
QmovedThermal energy the active option transfers across the envelope (rejected in heat, injected in cold) — capacity-capped at the rated kW each hour; unmet hours are echoed rather than hiddenkWh§18
EelElectric energy the heat pump consumes = heat moved ÷ COP (eq 18.15) — the number the planned financial analysis prices against the yield it protectskWh§18
SOHRemaining capacity as a fraction of new (1 = new, 0.8 = the end-of-life convention). The 25-year trajectory iterates it year by year with fade feeding back — year 1's usage recurring in absolute terms means every cycle on a faded pack is relatively deeper and every draw a relatively higher C-rate, so the cycle-damage term grows as SOH falls (the snowball-class accelerator; Attia et al. 2022 knee taxonomy). Plating/saturation knees are out of scope — no datasheet prints their basis—§18
SOCStored charge as a fraction of capacity (0 = the protection floor's reference empty, 1 = full). The battery model books it in energy terms; the BMS window (usable depth of discharge) bounds the cycled range—§18
OCVPack voltage at rest as a function of state of charge — the chemistry sets its SHAPE (LFP nearly flat mid-range, NMC sloped, lead-acid near-linear; per-cell anchor tables scaled to the pack's nominal voltage). Distinct from the module's V_oc (§9.3)V§18
RTEEnergy out ÷ energy in over a full cycle at the sheet's reference rate — the datasheet anchor the internal resistance is DERIVED from (eq 18.4), so the model reproduces the printed number exactly at its rating condition—§18
rvThe voltage-efficiency share of the printed round trip once coulombic losses are removed: r_v = RTE/η_c (eq 18.4) — the discharge-to-charge terminal-voltage ratio at the reference current—§18
R0Series (ohmic) resistance of the equivalent circuit — never a datasheet guess: derived from the printed round-trip efficiency at the reference rate (eq 18.4), then temperature-scaled by chemistry-class factors. Every rate- and temperature-dependence of loss flows through I·R₀Ω§18
VnomThe battery bank's nominal voltage (12.8/25.6/51.2 V lithium classes; 6/12 V lead-acid blocks) — sets the cell count that scales the per-cell voltage anchorsV§18
IrefCurrent at the chemistry's reference rate (0.5C lithium, the C10 rate for lead-acid — the conditions round-trip and capacity figures are printed at, NAMED in the response)A§18
IPack current at the operating point — charge or discharge; the I·R₀ product it forms is where every rate-dependent loss and heat term originatesA§18
IaccThe most current the pack can absorb at this state of charge and temperature (eq 18.6) — the voltage headroom between the charge-stage voltage and the rest voltage, over R₀. Constant-current/constant-voltage charging EMERGES from this: full headroom = the charger's current limit binds (CC); vanishing headroom = the taper (CV)A§18
VchgThe charger/BMS charge-stage (absorption) ceiling — universal per-cell conventions scaled to the pack: 3.65 V LFP, 4.20 V NMC, 2.40 V lead-acid per cellV§18
TbatThe pack's own temperature from the self-heating node (eq 18.10) — resistance, acceptance, the lithium 0 °C charge block and aging all evaluate at T_bat, never at raw ambient°C§18
QeffCapacity available at the actual discharge rate (eq 18.7) — Peukert scaling from the reference rate, times the chemistry's temperature class factor; clipped to 0.5–1.05× rated (beyond that the class tables would extrapolate, named)kWh§18
QratedNameplate energy capacity at the reference rate and 25 °C — the datasheet figure the capacity ladder resolves (kWh, or Ah × nominal voltage)kWh§18
nPeuRate–capacity exponent (eq 18.7): ~1.25 lead-acid (published 1.2–1.5 — a first-order effect there), ~1.02 lithium (near-flat). Named after Peukert (1897); the classic reason a lead-acid bank shrinks under heavy load—§18
FcycFraction of cycle life the year's profile consumed (eq 18.8, Miner's rule over the rainflow spectrum) — every stress factor equals 1 at the sheet's own rating condition, so a rating-condition cycle costs exactly 1/N_cyc (identity anchor)—§18
ciCount of the i-th extracted cycle in the rainflow decomposition of the year's state-of-charge series (1 full, ½ half — residuals count as halves)—§18
diDepth of discharge of the i-th rainflow cycle — a 100%-to-20% swing is d = 0.8; depths are stressed relative to the sheet's rated depth via the chemistry exponent k_p—§18
DoDrThe depth of discharge the printed cycle-life figure is quoted at (the pair travels together: 6000 cycles @ 80% ≠ 6000 @ 100%)—§18
kpChemistry class of depth sensitivity (eq 18.8): 1 LFP (near throughput-neutral), 2 NMC (deep cycles cost ~4× their share), 1.25 lead-acid — chemistry-review classes (LFP near-neutral per Wang et al. 2011's throughput law; the ~10× full-vs-shallow damage ratio for NMC from cycling reviews; lead-acid Ah-throughput weighting practice)—§18
NcycThe sheet's printed cycle count at DoD_r to the 80% end-of-life convention — the aging model's anchor point; chemistry-class defaults (6000/3000/1200) fill only when unprinted, provenance echoedcycles§18
frateThroughput-weighted mean of e^(k_C·C_rate) over the year, normalised to the reference rate (eq 18.9) — high-rate hours cost MORE than their energy share (the grid-coupling surface); ≡ 1 for an all-at-reference-rate profile—§18
ftempThroughput-weighted Arrhenius factor of the year's pack-temperature trace, normalised to 25 °C (eq 18.9) — hot hours age the pack disproportionately, and the trace is the pack's own temperature (eq 18.10), not ambient—§18
kCExponential rate-stress slope: from Wang et al. 2011's graphite-LFP throughput law (≈ +16% fade per +1C at 25 °C); the NMC class runs about twice as rate-sensitive (named-typical); lead-acid's rate physics is carried by the Peukert side instead—§18
CrateCurrent (or power) as a multiple of the one-hour-discharge rate: 1C empties the pack in one hour, 0.5C in two. The universal battery rate unit1/h§18
CrefThe chemistry's printed-figure reference rate: 0.5C lithium, 0.1C (the C10 convention) lead-acid1/h§18
kcalSquare-root-of-time calendar-fade scale (eq 18.9; Naumann et al. 2018 kinetics), stressed by the profile's mean state of charge and temperature; chemistry classes 1.4/2.2/3.0 %/√yr (LFP/NMC/lead-acid), named-typical%/√yr§18
tyrElapsed calendar time in years — calendar fade grows with its square root (diffusion-limited interface growth), so the second year costs less than the firstyr§18
CthHeat capacity of the pack = specific heat × mass (lithium cells ~1000–1100 J/(kg·K), Battery Design measured compilations; the datasheet weight wins the mass, else chemistry-class kg/kWh fills, provenance echoed)J/K§18
Q̇lossThe pack's own Joule heat — exactly the I²R₀ loss the energy ledger already books each hour (the Bernardi irreversible term). The entropic (reversible) term flips sign with state of charge AND current direction and cancels over any full cycle, so it is neglected on chemistry-independent grounds (named)W§18
hASurface conductance × area between pack and its surroundings — natural-convection cabinet class (≈ τ of 1.5 h at 10 kWh); sets WHERE the temperature settles (steady state = T_env + Q̇/hA) while C_th sets how fastW/K§18
TenvThe pack's surroundings: raw ambient, or the enclosure interior when one is configured (the §18 sol-air chain) — the temperature the self-heating node relaxes toward°C§18
TsaThe equivalent outdoor temperature a sunlit surface actually experiences (eq 18.12, ASHRAE construction): ambient plus absorbed sunlight over the exterior film coefficient, minus the longwave sky correction — a dark container face in full sun can 'feel' 40 °C above the air°C§18
GfaceSunlight on one enclosure face — each wall (90° tilt, N/E/S/W) and the roof (0°) gets its OWN hourly series through the same Hay–Davies transposition the panels use (§5), with paint treated as broadband (no glass optics)W/m²§18
hoCombined convection + radiation coefficient of the outside surface — the ASHRAE conservative standard 17 W/(m²·K). Also sets the bare-steel envelope U-value (films alone ≈ 5.6 W/(m²·K); foam classes add their d/k)W/(m²·K)§18
mpcmMass of phase-change material sized to hold the worst day passively (eq 18.13): excess heat above the setpoint over the latent heat of fusion. Only honest alongside the overnight regeneration margin — a melt point night air cannot reach never re-freezeskg§18
EexcThe worst single day's heat arriving above the cooling setpoint (25 °C, BESS practice) — envelope conduction under sol-air forcing plus the battery's own losseskWh§18
hfusHeat absorbed per kilogram of phase-change material melting at constant temperature — ~200 kJ/kg class (paraffin/salt-hydrate range 150–250). Written h_fus because L is reserved for latitude in this notationkJ/kg§18
PdcThe array's hourly DC power arriving at the inverter — the composed per-panel series after wiring effectskW§18
Pdc,0The inverter's rated DC input (its AC rating over nominal efficiency) — the part-load ratio's denominator; output above the AC rating clipskW§18
fpwmHourly harvest of a pulse-width-modulation (PWM)-coupled string as a fraction of the panels' available maximum (eq 18.2) — the string is pinned at the battery bus, so mismatch between the panels' best voltage and the bus is the whole loss (practice: MPPT harvests 5–30% more; COLD worsens PWM as the panels' voltage rises away from the bus). Zero when the bus exceeds what the string can reach — no charge flows at all—§18
VbusThe voltage a PWM-coupled string is pinned to — modelled at the midpoint of the bank's nominal and charge-stage voltages (the bulk-band convention, NAMED; factors at both bounds are echoed)V§18
IstrThe one series current a string carries — read off the string's summed current–voltage curve at the operating voltage (each panel's curve at its own shaded irradiance, §9.3 machinery)A§18
mstrHourly ratio of true string-wired power to per-panel-electronics power (eq 18.1) — exactly 1 when a string's panels see equal light (series MPP is additive over identical curves); below 1 under partial shade because every panel must carry the same current. Subscripted like m_sh: bare m is the air mass—§18
SeEffective plane-of-array irradiance S + b_bif(1−f_struct)·S_rear — drives cell temperature (§8) and power (§9) when the module is bifacial (IEC TS 60904-1-2 equivalent-irradiance method); the §9.3 curve model consumes it unchanged, so bifacial gain is a current gain by constructionW/m²§9.3, §17
bbifRear-side efficiency as a fraction of front-side — the lesser of the Isc and Pmax rear/front ratios at STC (IEC TS 60904-1-2, where datasheets print it as a bare φ; written b_bif here because φ is reserved exclusively for the azimuth family φs/φp — the same reservation that writes latitude as L). Typical: p-PERC 0.60–0.70, n-PERT 0.75–0.95, HJT above 0.90 (IEA PVPS T13-14:2021)—§17
fstructFraction of rear-incident light blocked by the mounting structure — default 0.05 (neither mount carries a torque tube; open racking sits at the low end); published spread 2–20% by structure type (IEA PVPS T13-14:2021)—§17
SrearIrradiance on the module's back face: rear sky diffuse + rear beam (fixed tilt only) + self-shading-corrected ground reflectionW/m²§17
ĪgndView-factor-weighted mean of the per-segment ground irradiance I_i — 120 segments over ±15 slant lengths, each losing beam over its shaded fraction and sky where the module masks it (view factors F_i→sky, F_p→i); analytic far field beyondW/m²§17
IiIrradiance reaching ground segment i: unshaded beam DNI·sin β·(1 − shaded_i) plus the sky diffuse the module does not mask, I_d·F_i→sky; the view-factor-weighted mean over all segments is Ī_gndW/m²§17
shadediFraction of ground segment i inside the module's beam shadow that hour — each module edge projected along the sun direction onto the ground; for the tracker the shadow always falls on the side the rear faces—§17
Fi→skyView factor from ground segment i to the unobstructed sky = 1 − segment→module view factor (the module is the only obstruction in the 2-D cross-section)—§17
Fp→iView factor from the module's rear face to ground segment i — the reciprocity form F_i→p · w_i/ℓ of eq 17.1, weighting each I_i into Ī_gnd; segments plus the analytic far field close exactly to (1 + cos Σ)/2—§17
Fi→pView factor from ground segment i to the module — Hottel's crossed-strings form, eq 17.1: (crossed − uncrossed string sums) over twice the segment width—§17
dcrossSum of the two crossed 'string' distances joining the ground segment's endpoints to the module's opposite edges (Hottel's crossed-strings construction, eq 17.1)m§17
duncrossSum of the two uncrossed 'string' distances joining the segment's endpoints to the module's near edges — crossed minus uncrossed, over 2·w_i, is the exchange view factor (eq 17.1)m§17
wiWidth of ground segment i — 120 segments spanning ±15 slant lengths either side of the modulem§17
ℓCross-section slant length of the module, fixed at 2.0 m (named simplification — the rear gain depends on the clearance-to-slant ratio only weakly over the 1.7–2.3 m module band). The module's edges sit at heights E and E + ℓ sin Σ, which is how clearance E enters every view factorm§17
fvisFraction of an hour the sun spends above the horizon line h(φ) — 4 substeps of 15 minutes, so the hour the sun crosses the skyline goes fractional instead of all-or-nothing (crossing-fraction refinement of the PVsyst far-shadings gate, K = 4). A fully clear hour keeps exactly 1, a fully blocked one exactly 0. Applied once to the shared direct normal irradiance, so every beam term — both fronts, the tilt sweep, the bifacial rear — sees the same gate—§4
fhorCosine-weighted fraction of the plane-visible sky dome that survives the horizon — constant per plane orientation because the skyline does not move with the sun (PVsyst's treatment; Quaschning & Hanitsch 1998), so it is computed once per geometry and multiplies the isotropic diffuse alongside F_d. The tracker interpolates its hourly orientations from a precomputed tilt×azimuth grid; a horizontal plane under a flat horizon at elevation e gives exactly cos²e—§4
fnearCosine-weighted fraction of the plane-visible sky dome surviving the near objects' silhouettes — an exact ray/rectangle mask on the same dome as f_hor (no polygon approximation), constant per orientation, tracker interpolated from the same grid. Folds into the isotropic diffuse multiplicatively with f_hor — conservative where the horizon and the objects mask overlapping sky—§4
hHorizon elevation above the astronomical horizon as a function of compass azimuth h(φ) — the skyline the sun is compared against, interpolated piecewise-linearly with wrap-around from the supplied points. A profile peaking under 2° is insignificant (PVsyst threshold) and resolves to no shading at all°§4
fworstPer-substring shade fraction at its worst cell strip — the strip most covered by the projected shadow that hour, from exact projection of each obstruction onto the module plane clipped against the strip grid. Crystalline-silicon bypass diodes respond to the worst cell, not the average, so this fraction drives the diode decision (worst-cell-limited doctrine)—§4
fmeanArea-proportional beam shade fraction of the whole module — the mean over its cell strips. Thin-film modules (monolithic construction, no substring knock-out) lose beam in proportion to shaded area, so this is their loss path exactly—§4
SkPlane-of-array irradiance seen by substring k in the exact bypass-diode solve: diffuse + ground + (1 − f_worst,k)·beam. Each substring's single-diode curve is evaluated at its own S_k; the module characteristic is their series sum with each substring floored at its diode conduction voltageW/m²§4
mshPer-hour ratio of shaded to unshaded power from the electrical tier (exact curve solve, or the partition / thin-film closed forms) — scales hourly power on both configurations while the linear datasheet path keeps driving the energy level. Exactly 1 at night; subscripted because bare m is the air mass (§1/§2)—§4
SiThe i-th low-light anchor irradiance — whichever of {200, 400, 600, 800} W/m² the datasheet characterises (or the technology-typical fallback, the same four points), plus the definitional 1000 W/m² (η_rel = 1 at STC); consecutive anchors S_i, S_i₊₁ bound each interpolation segmentW/m²§9
ILLight-generated current — scales with irradiance (eq 9.5); one of the five extracted parametersA§9.3
I0Dark current of the junction diode — climbs steeply with temperature through the bandgap (eq 9.5), dragging voltage down on hot daysA§9.3
RsContact and bulk resistance — drops voltage under load and rounds the MPP knee; constant with operating conditions (De Soto 2006)Ω§9.3
RshLeakage paths across the junction — bleeds current, shaping the low-light roll-off; refined toward the datasheet IEC 61853-1 low-light points when suppliedΩ§9.3
aModified ideality a = n·N_s·k_B·T/q — the diode knee sharpness in volts; one of the five extracted parameters, scales with temperatureV§9.3
nJunction quality factor (1 = ideal diffusion diode, 2 = recombination-dominated); the relaxed fit searches the physical 0.8–2.5 range. Within §9.3, n is this — not the §1 day count—§9.3
NsSeries cell count per module — datasheet; beware half-cut cells: 144 half-cells = 72 cells wired in series—§9.3
EgBandgap of the cell semiconductor — sets how steeply I_0 climbs with temperature; per technology: c-Si 1.121 eV (De Soto 2006), CdTe 1.5, CIGS 1.15eV§9.3
Eg,refBandgap at the 25 °C reference — the anchor of the linear E_g(T) slope in eq 9.5eV§9.3
VocDatasheet STC open-circuit voltage — extraction constraint; the cold-side design ceiling of the voltage windowV§9.3
IscDatasheet STC short-circuit current — extraction constraint; approximately the photocurrent at STCA§9.3
VmpDatasheet STC rated voltage at the maximum power point — exact in the exact fit, approximate (delta reported) in the relaxed fitV§9.3
ImpDatasheet STC rated current at the maximum power point — exact in the exact fit, approximate (delta reported) in the relaxed fitA§9.3
PmpModule maximum power = V_mp × I_mp — matched exactly in both fit tiersW§9.3
NsubEqual cell groups, one per bypass diode (= the datasheet diode count; 0 → one substring); must divide N_s evenly—§9.3
IModule terminal current — within §9.3, I is current [A], not an irradiance (the §2–§4 family)A§9.3
VModule terminal voltage [V] — the curve's independent variableV§9.3
ncapDays for soiling to reach the S_max ceiling = S_max / SRdays§10
tclearDays for one snow event to clear naturally = slide_hours·f_T / (8·avail) — hours of sliding conditions needed, divided by the sliding-capable hours each day supplies (h ÷ h/day = days)days§11
d⃗3-D hub-to-hub offset from receiver to neighbour pedestal (terrain heights included) — written d⃗ to stay distinct from the §10 cleaning interval d and the dΩ differentialsm§4
ŝUnit vector pointing at the sun (east/north/up frame) — shadows travel along −ŝ, which is also why the per-unit sweep's sunward→leeward order is exact—§4
n̂Unit normal of the receiver plane; ŝ·n̂ > 0 means the plane faces the sun, and d⃗·n̂ > 0 marks a neighbour as sunward (able to shade)—§4
AFArrhenius acceleration factor—§12
AOD550Aerosol optical depth at 550 nm—§10
availFraction of the 8 h/day reference sliding window the month actually supplies, inferred from its own yield: clip(monthly yield / 40 kWh/m², 0.15, 1). Dimensionless — this month's melt-driving sun as a fraction of a full-window reference month's; halving avail doubles the days a natural clear takes (§11)—§11
bonusView-factor estimate of the snow-brightening (ρ lifted to ρ_snow in gated months) that the hourly simulation already embeds in every surface's yield — reported to attribute the effect, never added on topkWh/m²§11
CFEnergy delivered as a share of running at rated output every hour = E_annual / (η·87.6)%§14
coverFraction of the month the panel is snow-covered; monthly loss = yield × cover—§11
cvtotalClimate and snow variability combined in quadrature (RSS)—§16
cvclimateCV from the 39-year NASA POWER joint simulation (irradiance, temperature, rain, wind, AOD)—§16
cvsnowCV of annual snow-loss fractions from the 19-year ERA5 record (2004–2022)—§16
CVRelative spread of annual yields = σ / μ (sample, ddof = 1)—§16
DNIDirect normal irradianceW/m²§2, §5–§6
EoTEquation of timemin§1
EannualAnnual energy yield per unit area — hourly W/m² summed over the year ÷ 1000kWh/m²§14
E2axTwo-axis tracking annual yield, computed every runkWh/m²§14
EftFixed-tilt reference annual yield, computed every runkWh/m²§14
elevsiteActual site elevation — Open-Topo-Data ASTER (~30 m)m§3
elevgridReference elevation — PVGIS grid-cell metadata (0 m on the clear-sky path)m§3
fuserUser dust / soiling multiplier—§10
fTTemperature factor for snow–glass bonding — peaks near freezing, where daytime melt refreezes into an ice bond (step ladder, Fig. 11.1)—§11
FdMartín–Ruiz angular loss integrated over the isotropic sky dome at tilt Σ — sky light arrives at every angle, so no single-θ IAM applies; always < 1 (≈ 0.95 for clean glass); shares the beam's α_r, so soiling lowers it too—§5, §6
FaMartín–Ruiz angular loss integrated over the ground band at tilt Σ — the band arrives near-grazing, so it loses most (≈ 0.80 at Σ = 30°); → 0 as Σ → 0, where the ground view factor vanishes too—§5, §6
AiHay–Davies anisotropy index — DNI ÷ E₀, the extraterrestrial normal irradiance (Spencer form, solar constant 1366.1 W/m²): the fraction of sky-diffuse light treated as circumsolar, arriving from the sun's direction. Zero under overcast (the model is then purely isotropic — an anchored identity), approaching ~0.8 under the clearest skies. Computed once per simulation from the horizon-gated DNI, so a terrain-blocked sun carries no circumsolar. Not the ASHRAE A (apparent extraterrestrial flux, §2) — different quantity, different letter case pattern—§5, §6
RbProjects sun-direction light onto the panel plane: max(cos θ, 0) ÷ cos θz — for a sun-tracking plane cos θ = 1 so R_b = 1/sin β. The denominator is floored at cos 89° so the ratio cannot spike when the sun grazes the horizon (the same guard the pvlib reference implementation carries). Multiplies the circumsolar share I_d·A_i; a sun behind the panel gives R_b = 0 — no circumsolar—§5, §6
TdGeometry term of the F_d integral: sin Σ + (π − Σ − sin Σ)/(1 + cos Σ), tilt in radians—§5, §6
TaGeometry term of the F_a integral: sin Σ + (Σ − sin Σ)/(1 − cos Σ), tilt in radians—§5, §6
gainTwo-axis yield advantage over the fixed-tilt reference = (E_2ax − E_ft) / E_ft × 100%§14
GHIGlobal horizontal irradianceW/m²§2
IAMIncidence angle modifier—§6
IbeamDirect-beam component on the panel plane (two-axis: = DNI; fixed-tilt: DNI·cos θ·IAM)W/m²§5, §6
IdiffuseIsotropic-remainder sky diffuse on the panel plane = I_d × (1 − A_i) × (1 + cos Σ)/2 — the circumsolar share I_d·A_i·R_b rides the beam component (Hay–Davies, §5)W/m²§5, §6
IgroundGround-reflected component on the panel plane = I_r × (1 − cos Σ)/2W/m²§5, §6
ItotalSum of beam + diffuse + ground-reflected on the panel plane (= S in §8/§9)W/m²§5, §6
IdDiffuse horizontal irradianceW/m²§2, §5–§6
IrGround-reflected irradianceW/m²§2, §5–§6
netFixed-tilt yield after the combined soiling loss = yield(α_r,eff) × (1 − s̄/100) — angular penalty and blockage without double countingkWh/m²§10
NOCTDatasheet nominal operating cell temperature (800 W/m², 20 °C ambient, 1 m/s wind) — the module form converts it to U₀ = 800/(NOCT − 20) − U₁; guarded 33–58 °C°C§8
P90Yield exceeded in 90% of years (1-in-10 bad year) = P50 × (1 − 1.28·CV)kWh/m²§16
P75Yield exceeded in 75% of years = P50 × (1 − 0.674·CV)kWh/m²§16
P50Typical-year (median) yield — the live PVGIS TMY simulation resultkWh/m²§16
SRSoiling rate%/day§10
s̄Mean soiling loss over one accumulate-then-reset cycle — peak/2 (triangular) until the peak hits S_max, then the flat-top trapezoid form of (9.2)%§10
sgnSignum function: +1 for a positive argument, −1 for a negative one, 0 at zero — in (1.7) it copies the hour angle's sign onto the azimuth (mornings negative/east, afternoons positive/west)—§1
slidehoursSnow slide-off time = 1 / (0.197·sin Σ)h§11
tlocalLocal clock time (mid-hour convention; TMY path shifts from UTC first)h§1
tsolarTrue solar time — the sun crosses the local meridian at exactly 12:00h§1
TambAmbient air temperature°C§8, §3
TcellModule cell temperature (K in the Arrhenius model)°C§8, §9, §12, §13
Tcell,FaimanCell temperature the Faiman model (§8) would give before the 25 °C active-cooling cap is applied°C§13
TeqArrhenius equivalent temperatureK§12
TiHourly cell temperature from §8, in kelvin (clipped to [230, 420] K to prevent numerical overflow)K§12
TPVGISPVGIS TMY air temperature°C§3
vwWind speed (at 10 m)m/s§8
yieldAnnual energy yield per unit areakWh/m²§10, §11
rateAnnual degradation rate%/yr§12
YnAnnual yield in year n of the 25-year projectionkWh/m²§12
Y0Year-zero annual yield entering the projection, before first-year degradation and the annual rate — the §14 annual yieldkWh/m²§12
AEffective top-of-atmosphere solar intensity, varying with Earth–Sun distance; looked up from the ASHRAE Handbook monthly tableW/m²§2
BatmTotal optical depth per unit air mass — all atmospheric constituents (Rayleigh, water vapour, ozone, aerosols), broadband. Not the aerosol-only AOD550 of §10, and no longer overloading the §1 day angle B—§2
CDiffuse irradiance as a fraction of DNI; looked up from the ASHRAE Handbook monthly table—§2
BDay angle°§1
dEffective cleaning interval for a scenario — scheduled visits and rain cleaning combined (§10 scenario table)days§10
IManual snow-clearing interval per scenario — ∞ (no action) / 30 / 7 / 1 / 0 (perfect)days§11
NNumber of hourly cell-temperature samples in the year's history (8,760)—§12
pSoiling level reached just before a cleaning reset = SR·d%§10
daysCalendar days in the month being evaluateddays§11
HHour angle (negative mornings, positive afternoons)°§1
LSite latitude, north positive (Masters 2013 / thesis convention — see notation note in §1)°§1, §6, §7
nrainEffective natural cleaning events = R/10 when R ≥ 5 mm/month (Kimber)—§10
neventsMonthly snowfall / 5 cm water-equivalent per event—§11
mRelative air mass (Kasten–Young)—§1, §2
nDay of year1–365§1
PPV power output per unit areaW/m²§7, §9, §14
SPlane-of-array (POA) irradiance on the cell — the same quantity as Itotal (§5/§6), written S in Faiman's (2008) notationW/m²§8, §9
PcurveModule power evaluated on the full single-diode I–V curve at a given per-substring irradiance pattern — the exact tier compares the shaded pattern against uniform lightW§4
fjShaded-area fraction of cell-strip j for the hour — an exact projected-polygon overlap, computed before any electrical response is applied—§4
E0Irradiance on a sun-normal plane above the atmosphere for the day (solar constant 1367 W/m² with the Earth–Sun distance correction) — the Hay–Davies anisotropy index measures how clear the sky is by comparing DNI against itW/m²§5, §6
tThe simulation's hour-of-year index (1…8760) — every hourly series is a function of t; sub-hour refinements sample instants t_k inside each hourh§4, §7, §14, §18
KNumber of instants sampled inside each hour by the crossing-fraction refinement — K = 4, the value the PVsyst comparison validated—§4
kIndexes a module's bypass-diode substrings in the per-strip shade aggregation — and the sub-hour sample instants t_k in eq 4.1—§4
jIndexes the cell-strips within a substring — each strip gets its own exact shaded-area fraction f_j—§4
CHow many cell-strips a substring carries — the denominator of its mean shade fraction—§4
iIndexes the N = 8760 hourly cell temperatures inside the Arrhenius average (eq 12.1) — hot hours dominate the exponential mean—§12
nYear of operation in the 25-year projection (eq 12.4) — year 1 applies the one-off LID step, later years compound the annual Arrhenius rate—§12
iIndexes the ground strips between the rows — each is sunlit or shaded per hour, and the rear side integrates their view-factor-weighted irradiance (eq 17.2)—§17
iIndexes the rainflow-extracted charge/discharge cycles in the Miner damage sum (eq 18.8) — each cycle contributes by its depth relative to the rated convention—§18
kYear counter in the multi-year state-of-health recursion (eq 18.11) — each year's fade is computed on the already-faded pack, which is what makes degradation compound—§18

Constants & fixed parameters

SymbolMeaning (value)UnitsUsed in
Ea,cyc31.7 kJ/mol (Wang et al. 2011) — the Arrhenius energy scale of cycle aging. Distinct from the §12 module-degradation E_a (0.6 eV): different physics, different units, deliberately separate rowsJ/mol§18
Rgas8.314 J/(mol·K) — pairs with E_a,cyc in the aging Arrhenius factor (eq 18.9)J/(mol·K)§18
SrefStandard-test-conditions irradiance (1000 W/m²) — the reference of the eq 9.5 current scalingW/m²§9.3
VfForward voltage clamping a bypassed substring at −0.6 V (Schottky; Fronius engineering whitepaper, MDPI Energies 2021, 14:4778); dormant under uniform lightV§9.3
qElementary charge (1.602×10⁻¹⁹ C) — with k_B in J/K, k_B·T/q = 25.7 mV per cell at STCC§9.3
U0Faiman zero-wind heat-loss coefficient (default 25.0, c-Si open rack) — per-module overridable, or derived from the datasheet NOCTW/(m²·K)§8
U1Faiman wind heat-loss coefficient (default 6.84) — W/(m²·K) per m/s of wind; per-module overridableW·s/(m³·K)§8
c1First fitting constant of the integrated angular-loss factors: 4/(3π) ≈ 0.4244 (Martín & Ruiz 2001)—§5, §6
c2Second fitting parameter of the integrated angular-loss factors: 0.5·α_r − 0.154 (linear form per IEC 61853-3)—§5, §6
KsoilSoiling rate coefficient (0.65)%·day⁻¹·AOD⁻¹§10
Eac-Si degradation activation energy (0.6)eV§12
TrefReference temperature (298.15 K = 25 °C) — the STC anchor of the §9.3 translation and the Arrhenius reference in §12K§9.3, §12
kBoltzmann constant — 8.617×10⁻⁵ eV/K with the bandgap in eV (eq 9.5, §12); equivalently 1.381×10⁻²³ J/K, where k_B·T/q = 25.7 mV per cell at STC (eq 9.3)eV/K · J/K§9.3, §12
LIDYear-1 output drop vs nameplate — light-induced degradation (LID) plus any other year-1 effect; datasheet figure → per-technology warranty-class ceiling → 2% default%§12
ΓICAO standard atmospheric lapse rate (0.0065 °C per metre ≈ 1 °C per 154 m)°C/m§3
SmaxSoiling saturation ceiling — losses stop growing at 30% (Ferrada et al. 2018)%§10
kslideFraction of panel slant height cleared per hour of sliding (Marion et al. 2013)h⁻¹§11
rbaseBaseline degradation rate at 25 °C — resolution: datasheet warranted rate → per-technology median → 0.5 %/yr c-Si default (Jordan & Kurtz 2013/2016)%/yr§12
z1.28 = 10th percentile (P90), 0.674 = 25th percentile (P75) of the standard normal distribution—§16
dΩAn infinitesimal patch of sky in the view-factor integrals — dΩ = cos α dα dφ, the cos α keeping a patch's true area as it nears the zenithsr§4
dLeibniz's 'infinitesimal step in' — dP/dV is the slope of power against voltage, dT/dt the rate of temperature change, dΩ a sliver of sky—§4, §9, §18
eBase of the natural exponential — the Martín–Ruiz reflection factors, the battery stress factors and the ground-temperature wave are all exponential laws≈ 2.71828§5, §6, §18
πThe circle constant — 2π/8760 converts the hour of year into the angular phase of the annual ground-temperature wave (eq 18.16)≈ 3.14159§18