Why your solar forecast is wrong, and how forecasts are made
A solar forecast is a weather forecast with extra steps, and each step adds error. Knowing where the error comes from tells you how far to trust the number, and how to run a battery on a forecast that will sometimes be wrong.
What a solar forecast actually is
Nobody forecasts your panels directly. A solar forecast is built in three stages, and only the last knows about your roof.
- A weather model. A numerical weather model divides the atmosphere into a grid of cells and steps the physics forward from the latest observations: pressure, temperature, humidity, wind, and the water in each cell as vapour, droplets and ice. It runs a few times a day and produces, among much else, a cloud forecast for every cell and every hour.
- Irradiance. From the cloud, the water vapour and the sun's position, the model estimates how much sunlight reaches the ground in each cell, split into the direct beam and the diffuse light scattered by the sky.
- Your array. The irradiance is then projected onto your panels using their size, tilt and orientation, and turned into kilowatts with an allowance for panel temperature and the inverter.
Every solar forecasting service works this way in outline, differing in which weather models it uses and how it treats the array.
Why cloud is the hard part
The sun's position is known to the second, so a clear-sky forecast for any place and hour is essentially exact. All of the difficulty is cloud.
Cloud is hard for a weather model for two reasons. The first is scale. A cumulus cloud is a few hundred metres across and lives for twenty minutes, while a model cell is a few kilometres wide and a model step is an hour. The model cannot see individual clouds; it estimates what fraction of a cell is covered and how thick the cover is, on average. On a day of broken cloud, that average can be right while your roof spends the hour alternately in full sun and deep shade.
The second is sensitivity. Thin high cloud cuts a little from the direct beam; a thick stratus deck cuts most of it. The difference between the two is a small change in the model's water content, and small errors in that produce large errors in irradiance. A day that the model calls "mostly cloudy" can produce half of a clear day's energy or a fifth of it, depending on cloud type, and the model is often unsure which.
Resolution and lead time
The grid
Forecast providers use models with cells from around a kilometre to tens of kilometres across. Finer cells help with coastlines, hills and valleys, where weather differs over short distances, but they still do not resolve individual clouds. Locating your site to the nearest street is more than enough; the forecast cannot tell your house from your neighbour's.
Why today's forecast beats last week's
Weather models are re-run several times a day from fresh observations, and each run corrects the last. Errors grow with the time between the run and the hour being forecast. A forecast for this afternoon made this morning is usually good; the same afternoon forecast a week ago was little more than climatology. The practical rule for a battery is to act on the latest forecast for the next day or two and ignore anything further out.
The array details that matter
The third stage is where you can remove error yourself, because the inputs are yours. Get these right and the forecast tracks your real output much better.
- kWp per array. The installer's paperwork lists the rated power of each string. If your roof has two aspects, enter them as two arrays, not one average.
- Tilt. A typical UK pitched roof is 30–40°. Flat-roof frames are often 10–15°. A few degrees of error is minor; treating a 15° roof as 35° is not.
- Orientation. The compass bearing the panels face: 180 is south, 90 east, 270 west. East and west arrays make roughly the same energy over a day as each other but at different hours, and a forecast that has them the wrong way round will be wrong every morning and every afternoon.
- Shading. A chimney, a tree or a neighbouring roof that shades part of the array at certain hours takes a bite the forecast cannot see. The bite moves with the season as the sun's path changes.
- Inverter clipping. If the array's peak power is higher than the inverter can pass, sunny middays are flattened at the inverter's limit. A 5 kWp array on a 3.68 kW inverter never shows more than 3.68 kW, whatever the forecast says. A Powerwall 3's built-in inverter accepts up to 20 kW of solar, so clipping there is rare; an AC-coupled Powerwall 2 depends on the separate solar inverter.
- Soiling and age. Dust, pollen and slow degradation take a few percent the rating ignores.
Learning a correction from your own history
Most of those array errors have one useful property: they are systematic. A wrong tilt, a shaded corner or an ageing array all reduce output by a similar proportion on similar days. That means the forecast can be corrected by comparing what it predicted with what the meter recorded, day after day, and scaling future forecasts by the ratio.
After a week or two of metered data a single learned factor removes most of the bias. If your forecasts are consistently 15 % high, the correction pulls them down by 15 %, and the remaining error is the weather's, not yours. This is why a forecast that starts out poor settles down over a fortnight, and why it should be reset if you change the arrays. It cannot fix cloud; nothing can. It fixes the part that was your roof.
How to use a forecast that will sometimes be wrong
A forecast is not a promise. Used well, it is a range with a most-likely value, and you plan for the range.
- Plan on the pessimistic side for anything expensive to get wrong. Deciding whether to buy cheap overnight energy is the classic case: if the forecast says 15 kWh tomorrow, treat it as "probably at least 10" and buy enough overnight to cover the shortfall. Being slightly over-charged on a sunny day costs a little export; being empty on a grey day costs peak-rate import.
- React during the day. By late morning the day has largely declared itself. If the panels are well ahead of forecast, there is room to lower the reserve; if they are well behind, hold more back for the evening. The correction is made while there is still time to act. Overnight buying decisions are covered in Grid charging: when it pays and when it doesn't.
- Look at the shape, not just the total. A day of morning cloud and a clear afternoon fills the battery late; the same total spread evenly fills it by lunch. An evening plan depends on which.
What "solar remaining today" means
The most useful single number for a battery is not today's total but how much of it is still to come. At 14:00, "solar remaining today" is the forecast for the hours between now and sunset. It answers the question that matters in the afternoon: will the panels fill the battery before dark, or should the battery be spared? A rule such as "if solar remaining today falls below 5 kWh and the battery is under 40 %, raise the reserve" uses the forecast as it deserves to be used, with the morning's uncertainty already resolved.
In SunHarvest
SunHarvest takes hourly radiation from the Open-Meteo weather service and runs it through the arrays you enter in Settings › Site, with kWp, tilt and bearing; the forecast reaches two days ahead and refreshes hourly. After a week or so of metered data a learned factor scales it to what your panels actually produce. The Energy page draws each past day's forecast dashed beside what was recorded, so you can judge it yourself. Rules and the AI planner read the same forecast through signals such as solar remaining today and solar tomorrow, and on a Powerwall 3 the Android app's local gateway reads each string separately, the quickest way to spot a shaded array. See the site and forecast guide.
General information, not financial or electrical advice. Tariff terms, prices, warranties and connection rules change and differ by supplier, region and installation; check the current documents for yours. SunHarvest is not affiliated with Tesla or Octopus Energy.