Solar irradiance describes the sunlight arriving at a surface at a particular moment. It is an input to a photovoltaic power forecast, not the power forecast itself. For a trader, confusing the two can mean comparing forecasts that describe different physical quantities. For an operator, it can conceal a plant issue behind a weather explanation.
GHI, DNI and DHI describe different surfaces
Global horizontal irradiance, or GHI, is the total solar radiation reaching a horizontal surface. Direct normal irradiance, or DNI, describes the direct beam on a surface facing the sun. Diffuse horizontal irradiance, or DHI, describes radiation scattered by the atmosphere and clouds onto a horizontal surface. Their geometric relationship is GHI = DHI + DNI × cos(z), where z is the solar zenith angle. The angle is measured from vertical, not from the horizon.
Irradiance is normally expressed in watts per square metre. Irradiation is the energy received over an interval, often expressed in kilowatt-hours per square metre. A file containing an instantaneous W/m² value cannot be treated as an interval energy total without accounting for time. Ask whether a forecast value is instantaneous, an interval average or an accumulation. This distinction becomes especially important when converting weather-model radiation fields.
From sunlight to electricity
PV modules rarely lie horizontally. A model first estimates the radiation on the plane of the array, using the direct and diffuse components, sun position, surface orientation and ground reflection. Fixed arrays and trackers therefore need different geometric information. The power model then accounts for module behaviour, temperature, conversion to AC power and the relevant system losses or constraints.
Two nearby systems can receive similar horizontal irradiance but produce different power because they have different tilt, orientation, tracker operation, installed DC capacity or inverter limits. A forecast that gets the weather right can still get delivered electricity wrong. Keep the irradiance and power stages separate when investigating errors, rather than asking one aggregate score to explain the whole pipeline.
How irradiance forecasts are made
Numerical weather prediction estimates future atmospheric conditions and radiation. Satellite observations can help describe cloud fields and their motion, while local sensors can reveal conditions at the plant. Statistical methods can correct recurring differences between an input forecast and observations. The useful balance depends on the decision horizon: recent cloud information matters differently for the next delivery interval than for tomorrow's production profile.
Cloud development remains a major source of uncertainty. A model may predict broadly correct cloud cover while mistiming a cloud edge at one plant. Dust, aerosols and snow can add other difficulties. An attractive smooth curve does not show whether these effects are represented. Request validation during variable-cloud periods, not just examples from clear days.
Check the measurements before scoring
A sensor is evidence, but it is not automatically a perfect reference. Orientation, calibration, cleanliness and shading can change its readings. A horizontal sensor should not be compared directly with a tilted-array forecast. Timestamp conventions also matter: a value labelled noon may represent the preceding interval in one system and the following interval in another.
Start an evaluation with a data dictionary. Record the measured quantity, units, sensor plane, interval convention, timezone and quality flags. Preserve missing values rather than silently replacing them with zero. Then compare like with like, grouping results by daylight conditions and lead time. Keep nighttime handling explicit so a large number of zero readings does not disguise daytime errors.
Questions for a forecasting provider
Ask which radiation quantities can be delivered, how they are converted to plant output and what plant metadata is required. Request an explanation of uncertainty, forecast issue times and delayed-data behaviour. For a backtest, require forecasts that would actually have been available at each decision time. A retrospectively reconstructed weather series is useful for some analyses but is not automatically evidence of live forecasting performance.
For portfolio analysis, retain each site’s geometry and capacity rather than assuming one radiation-to-power relationship for all plants. Summing correctly modelled site production is a clearer starting point than converting a regional irradiance average into an unspecified regional power total.
The practical outcome is a traceable chain: weather input, array geometry, power conversion and operational availability. When that chain is documented, a poor production forecast becomes a problem that can be investigated. Without it, an irradiance score risks becoming a substitute for understanding the electricity your business needs to schedule.