Solar irradiance forecasting across horizons: information-source boundaries, hand-offs, and evaluation requirements

This is the authors' abstract. We don't add key points for this paper.

Solar irradiance forecasting supports renewable-energy integration because reserve allocation, storage dispatch, curtailment management, and distribution-level voltage control depend on forecasts whose errors can affect operational risk and cost. At intra-hour to intra-day lead times, cloud variability limits predictability, so performance depends not only on the final learning model but also on how cloud fields are observed, represented, evolved, and related to surface irradiance. This review provides a process-based reading structure for critically comparing these information pathways without proposing a universal architecture. It distinguishes local camera-based nowcasting, regional satellite nowcasting, and numerical-weather-prediction-driven intra-day and day-ahead forecasting, and identifies the conditions under which each source can serve as primary predictive information. Across these contexts, it examines observation and retrieval uncertainty, cloud evolution, radiative conversion, prediction, calibration, forecast combination, and decision-oriented verification. The central conclusion is conditional: current-image extrapolation is useful only while observed cloud organisation remains informative for the specified target, whereas numerical weather prediction and calibrated post-processing become primary when large-scale atmospheric evolution governs the forecast. The review therefore focuses on admissible information, failure modes, and reporting requirements—target variable, issue time, lead time, spatial support, reference forecast, regime, and verification design—needed to compare claims across studies and connect forecast quality to operational decisions.