Global Horizontal Irradiance (GHI) Forecasting

GHI is defined as the total shortwave radiant flux incident on a horizontal plane per unit area, measured in W/m². It is the sum of two components: Direct Normal Irradiance (DNI) projected onto the horizontal surface via the cosine of the solar zenith angle, and Diffuse Horizontal Irradiance (DHI) scattered by atmospheric constituents. This geometric constraint means a GHI forecast is not merely a cloud prediction; it is intrinsically linked to precise solar geometry calculations.

Unlike wind speed, GHI possesses a strong, deterministic diurnal cycle defined by the sun's position. The extraterrestrial irradiance and theoretical clear-sky irradiance provide a mathematically precise upper bound that collapses to zero at night. Forecasting models must learn the residual between this known envelope and the actual irradiance, often by predicting the Clear Sky Index to isolate stochastic cloud-driven attenuation from the deterministic astronomical signal.

GHI variability spans multiple temporal scales, requiring distinct modeling strategies:

• Intra-hour (0-60 min): Dominated by rapid irradiance ramp rates caused by moving cumulus clouds, demanding high-frequency sky imagery or Cloud Motion Vector (CMV) techniques.
• Intra-day (1-6 hours): Requires advection of satellite-observed cloud fields.
• Day-ahead (24-48 hours): Relies entirely on Numerical Weather Prediction (NWP) model output for cloud formation and dissipation physics.

The primary source of GHI forecast error is the misprediction of cloud optical depth and coverage. Key attenuating phenomena include:

• Cumulus humilis: Causes high-frequency, high-amplitude ramp events.
• Cirrostratus: Produces thin, uniform attenuation often missed by coarse NWP models.
• Aerosol optical depth: Dust, smoke, and pollution introduce systematic bias, particularly in arid regions.
• Soiling loss: A cumulative degradation factor on the photovoltaic panel itself, distinct from atmospheric attenuation, that must be modeled separately.

Point GHI forecasts for a single pyranometer are inherently more volatile than spatially aggregated forecasts for a utility-scale photovoltaic plant. As geographic area increases, high-frequency ramp rates are smoothed due to the spatial decorrelation of cloud shadows. A robust forecasting system must model this spatio-temporal averaging effect, often using Graph Neural Networks where nodes represent distributed irradiance sensors and edges represent spatial covariance learned from historical data.

Due to the chaotic nature of cloud physics, a single deterministic GHI value is insufficient for operational decision-making. Modern forecasting systems must output a full predictive distribution or quantile range. This is achieved through:

• Ensemble NWP: Running multiple physics parameterizations.
• Quantile Regression: Trained with Pinball Loss to directly estimate specific quantiles.
• Analog Ensemble (AnEn): Searching historical archives for similar atmospheric states. The resulting Probabilistic Power Forecast enables grid operators to dynamically size operating reserves based on quantified uncertainty.