Smoke dims the sky, not the glass

Soiling settles on the module; smoke attenuates sunlight in the atmosphere before it arrives, which is why its losses appear and vanish with the plume instead of accumulating. Part 05 covered the first kind of particulate; this part covers the second. The loss scales almost linearly: New South Wales rooftop fleet data puts it at about 13 per cent of PV energy per 100 micrograms per cubic metre of PM2.5, consistent with the 12.5 per cent insolation loss per 100 micrograms found across urban-haze cities globally. At the concentrations major fire events now produce (Sydney's Black Summer peaked near 350 micrograms per cubic metre, New York's 2023 Canadian-smoke day at 185, California's 2020 statewide peak at 659), that arithmetic gets severe fast.

The measured record, restricted to energy and output figures so unlike quantities are never averaged:

From the 18-event ledger in the public dataset. Each impact applies over its own window and must not be read as an annual loss. Irradiance-basis observations (Athens 2021: 10-20% daily GHI on fire days; southern Brazil 2024: 10-20% clear-sky irradiance) and blank-by-rule rows (NYISO published megawatts, not a percentage; no confirmed Jakarta yield figure) are recorded in the dataset with their bases flagged.

The pattern across the ledger is consistent: single days lose a quarter to nearly half of output, fortnights lose around an eighth, and even the archetypal sustained event, two months of Black Summer smoke across Australia's grid, cost 9.5 per cent of grid-scale solar. Large, but short-window. A 160-system NSW study put the whole 71-day episode at 4.2 per cent of energy, roughly 175 GWh state-wide.

Two opposite trends, and the one that matters

The global fire record runs in two directions at once. Satellite burned area fell roughly 24 to 27 per cent over two decades, and that decline is real; it is driven by savanna-to-cropland conversion in Africa and South America, grass fires with low smoke plumes far from any PV fleet. Meanwhile forest-fire carbon emissions rose 60 per cent between 2001 and 2023, human fire exposure rose 40 per cent even as area burned shrank, and Canada's 2023 season alone emitted about five times its 20-year average. Forest and boreal fires are the ones that loft dense smoke over cities and grids.

For solar, the relevant trend is the rising one, and it is regionally specific. The measured 20-year record supports treating smoke as a seasonal yield factor in three regions and only three: the North American West (California burned area up fivefold since the 1970s, critical fire-weather days up from 14 to 34 a year, the 2020 insolation deficit more than twice any year back to 2001), south-east Australia (forest burned area up 350 per cent between the periods 1988-2001 and 2002-2018 before Black Summer, with AEMO now modelling smoke explicitly in its solar forecasts), and Mediterranean Europe (significant fire-weather trends across most of Iberia's burnable area, and 2023 producing the largest fire ever recorded in the EU). South-East Asia's haze is recurring but El Nino-linked and trending down in severity; East Asia's chronic attenuation is anthropogenic aerosol, a different ledger; and for smoke-importing regions like the US Northeast, transported plumes set air-quality records without touching the long-run resource, which stays within about 5 per cent of baseline.

The resource survives smoke

The resource survives smoke: in every documented market the long-run mean irradiance loss is small against year-to-year weather noise, so smoke is not a reason to skip solar anywhere in the baseline. The planning numbers are another matter. In the three structural regions a P90 yield estimate should carry a fire-season term, because a Black Summer costs a summer 6 to 13 per cent of grid-scale output and a 2020-scale September cost California an eighth of a fortnight's energy. Storage sizing moves at the margin, since a battery specified for multi-day autonomy covers the smoke weeks when output runs 20 to 40 per cent under forecast for days. And the forecast is where the error concentrates: day-ahead solar forecasts overestimated Californian output by an average 27 per cent on heavy-smoke days, and grid operators on two continents now model smoke because of it. On that record, a yield model or day-ahead forecast in the three structural regions that carries no explicit smoke term is measurably optimistic. Owners in smoke regions should also rinse panels after ash-fall events; ash that settles is soiling, and part 05's rules apply.

Eighteen events, and how each was graded

The event ledger holds 18 market events across 12 of the 40 baseline markets, each with its window, measurement basis (grid operator, fleet, satellite model or study), confidence grade and source; ranges are recorded at midpoints and flagged. The ledger records events rather than a climatology: annualising it requires the regional frequency evidence, and that conversion is modelling work that sits outside the dataset. Irradiance-basis rows are flagged and never averaged with energy-basis rows. And where no credible smoke-on-PV quantification yet exists, as for the desert-dust and Sahel markets whose particulate story belongs to part 05, the ledger's cells are blank, with the reason recorded in place of a proxy.