Solar Forecasting When the Sky Is the Variable: Clouds, Smoke, Snow, and Dust

Solar forecasting is a fundamentally different problem from wind forecasting. The sun's position in the sky on any future date is known to within a few seconds. The seasonal arc of daylight hours is fixed. What is uncertain is what sits between the sun and the panel: clouds, smoke, dust, and snow. A solar forecast is, almost entirely, a forecast of the atmospheric column above the array. This article walks through how that forecast is built, why some sky conditions are harder than others, and what operators watch when the sky is doing something unusual.

Solar forecasts split cleanly by time horizon, and each horizon uses a different physical tool. Inside about thirty minutes, the dominant technique is sky-imager-based nowcasting. Ground-based or array-mounted cameras track cloud motion across the visible hemisphere and project the next several minutes of irradiance forward. This is what an individual plant operator uses to anticipate a passing cumulus shadow.

From thirty minutes out to roughly six hours, the dominant tool is satellite-derived irradiance combined with cloud-motion vectors. Geostationary satellites — GOES-East and GOES-West for North America — produce visible and infrared imagery every five minutes. From these images, providers derive cloud optical depth, cloud-top height, and cloud type, then translate those into surface irradiance using radiative-transfer models. Cloud motion vectors derived from successive images are extrapolated forward to project where the cloud field will be in the next several hours.

Beyond six hours, the dominant tool is numerical weather prediction. The HRRR (High-Resolution Rapid Refresh) provides hourly updated three-kilometer forecasts out to eighteen hours, with explicit cloud and radiation physics. The NAM, GFS, and ECMWF push the horizon out further, at coarser resolution. ISO solar forecasts blend these NWP outputs with satellite-derived nowcasts, statistical post-processing, and sometimes plant-level production data.

The accuracy curve across these horizons is steep and well-known. Inside thirty minutes, root-mean-square error on fleet-level irradiance can be below 5 percent of clear-sky for many sites. From thirty minutes to six hours, error climbs into the 10–20 percent range. Beyond a day, error is highly weather-pattern dependent — a clear-sky high-pressure ridge can be forecast within a few percent two days out, while a dynamic frontal pattern with embedded convection can be off by 30 percent or more.

A persistent overcast deck is operationally easy to forecast even if it is bad for production. Stratus is large-scale and slow-moving, NWP models resolve it well, and the irradiance reduction is roughly uniform. A clear-sky day is similarly easy: the only inputs are solar geometry and aerosol depth, both of which are well-characterized.

Convective clouds — fair-weather cumulus, building cumulus congestus, and full thunderstorms — are the hard case. They are small relative to model grid cells, they form, grow, and dissipate on timescales of tens of minutes, and they cast shadows that can cut plant output by 70 percent or more for a few minutes at a time. NWP models, even at three-kilometer resolution, do not resolve individual cumulus towers; they parameterize the convective field statistically, which works for daily totals but fails badly for sub-hourly variability.

The result is that on a typical summer afternoon with scattered convection, the day-ahead forecast can be excellent in total energy but poor in five-minute variability. Operators see this as a forecast that nails the integrated megawatt-hours but oscillates wildly around the actual five-minute curve. The economic consequence is real: solar generators are exposed to real-time price volatility on exactly the hours the forecast was least confident.

Wildfire smoke has emerged as a recurring solar-forecasting variable across the western U.S. and, increasingly, the eastern U.S. when Canadian smoke transports south on northerly flow. Smoke is both an atmospheric column problem and a panel-surface problem. In the column, smoke acts like a partial-cloud filter — heavy aerosol loading reduces direct-beam irradiance, the component that single-axis-tracking utility plants depend on most, by a larger fraction than it reduces diffuse irradiance. A typical heavy smoke day across CAISO can cut fleet output by 10–30 percent below a clear-sky baseline even when the sky looks merely hazy rather than overcast.

On panel surfaces, smoke deposits a film of fine particles. A few days of smoke followed by a dry stretch can leave residual soiling that is not cleaned by routine rainfall, particularly in arid climates. Plant performance ratios drop until the next manual wash or first significant precipitation. Most ISO solar forecasts do not yet ingest smoke-specific aerosol products in real time, so on smoke-affected days, the forecast is often biased high relative to actual output. Sophisticated trading desks watch the GOES aerosol optical depth product and the NOAA smoke plume forecasts to anticipate this bias.

Snow is a binary problem. A panel covered in snow produces essentially zero output regardless of solar geometry. A panel that has shed its snow produces near-design output. The forecasting question is when, and at what tilt and snow density, the snow will slide.

Empirical studies of utility-scale arrays in snowy climates suggest that snow tends to clear from fixed-tilt panels within one to three days of a storm if temperatures rise above freezing and the sun is available, faster on steeper tilts, slower on flatter installations. Single-axis trackers can be moved to a steep stow angle to encourage shedding. Two-axis trackers can be tilted nearly vertical, which clears them quickly but exposes the back of the array to wind. The forecasting challenge is not the storm itself — snow accumulation forecasts from NWS are reasonable — but the post-storm evolution. A clear, sunny day after a storm clears panels within hours. An overcast day with sub-freezing temperatures can leave them snow-covered for a week.

Operators watch the post-storm forecast more carefully than the storm forecast itself. A 12-inch snowfall followed by sustained sun and 40°F highs is operationally minor. The same snowfall followed by overcast skies and sub-20°F highs can take a multi-gigawatt fleet offline through the coldest hours of the following week.

Long-range dust transport is a less familiar variable for North American operators but matters seasonally. Saharan Air Layer events transport dust across the Atlantic in summer and can affect solar output in the southeastern U.S. and the Gulf Coast. Asian dust events occasionally reach the West Coast in spring. Volcanic stratospheric aerosol from major eruptions can affect global irradiance for months. The mechanism is the same as smoke: increased aerosol optical depth reduces direct-beam irradiance more than diffuse, and deposition on panels reduces performance ratios until cleaning.

Dust events are typically forecast a few days in advance through NOAA's HYSPLIT trajectory models, ECMWF's CAMS aerosol forecasts, and university research groups. They are slower-moving and longer-duration than convective cloud events, which makes them easier to incorporate into day-ahead solar forecasts once an operator knows to look for them.

Three practical habits help. First, look at the synoptic pattern, not just the irradiance number. A clear-sky high-pressure ridge gives you a tight forecast. A weak trough with embedded convection gives you a soft forecast even if the headline numbers look similar.

Second, watch the satellite imagery as the day progresses. Real-time GOES visible imagery shows you whether the actual cloud field is matching the forecast cloud field. If the model placed a clear slot over your fleet at noon and the satellite shows a thick cumulus deck, the rest of the day's forecast is suspect.

Third, ask whether anything unusual is in the column. Smoke plume forecasts, dust transport advisories, and post-snowstorm panel coverage are not in most standard solar forecasts. On days when any of these is present, the forecast is likely biased and the bias is usually toward over-prediction.

The wind and solar panel for each ISO surfaces the day-by-day forecast contribution from the solar fleet, which is useful for catching multi-day patterns driven by an extended ridge or a persistent overcast trough. The CPC 6–10 and 8–14 day temperature outlooks correlate strongly with insolation patterns at the seasonal scale, since temperature anomalies and synoptic regimes are not independent. The drought monitor and the hurricane and severe weather panels surface the smoke, dust, and convective conditions that bias short-term solar forecasts. Watching these alongside the city-level cloud and precipitation forecasts gives you a meaningful read on whether tomorrow's solar day is likely to come in close to forecast or well below it.

Concepts in this article reflect public-use materials from the federal agencies, federal labs, and operators listed below.