Résumé

This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.

• Combine different input data and forecasts (model blending); here, very short-term forecasting, up to approximately 1 hour ahead, greatly benefits from the use of local online irradiance or PV power measurements as input. • Derive PV power forecasts from meteorological forecasts.

Besides time-series forecasting, ML algorithms are increasingly used for image prediction using ASI or satellite data, e.g., to compute the optical flow in cloud motion approaches.

State-of-the-art PV power forecasting services do not rely on a single forecasting model but integrate different inputs and models. Prominent examples are intraday forecasting systems up to several hours ahead integrating online measurements, satellite-based forecasts, and NWP model forecasts or day-ahead forecasting systems combining different NWP models, both using statistical and/or ML algorithms for forecast optimization.

Besides forecasting for single PV power plants and portfolios of PV plants, the estimation and forecasting of regionally aggregated PV power is important for grid operators for marketing of PV power and grid management. Here, an additional challenge is that PV power is not measured at a sufficient resolution for most plants in many countries, and information on PV systems is incomplete. Still, because of spatial smoothing effects, forecast errors of regionally aggregated PV power as well as virtual power plants (VPPs) (normalized to their installed power) are much smaller than for single PV plants, depending on the size of the region and the set of PV plants contributing.

Forecast evaluations provide users with necessary information on forecast accuracy, assisting them in choosing between different forecasting services or assessing the risk when a forecast is used as a basis for decisions. Beyond general information on the overall accuracy of deterministic forecasts, probabilistic forecasts provide specific uncertainty information for each forecast value, depending on the weather conditions, and they allow for better risk management.

Chapter 9-3 This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications. Figure 9-1. Different forecasting methods suitable for various spatial and temporal scales Empirical and/or physical models are combined with statistical and/or ML models for forecast optimization. The spatial scales of the forecasting methods are defined by spatial resolution and spatial coverage. The temporal scales are defined by temporal resolution, update frequency, and forecast horizon.