A Radiative‐path‐aware Satellite‐Based Irradiance Retrieval Framework Supporting Ultra‐short‐term PV Power Forecasting

ABSTRACT

Accurate ultra‐short‐term photovoltaic (PV) forecasting is fundamentally constrained by a spatial‐scale mismatch: while irradiance dynamics evolve over wide areas, forecasts for isolated PV stations rely primarily on point‐scale measurements. To address this, we propose a paradigm shift that establishes scenario‐adaptive irradiance retrieval as a foundational data service. The core innovation is a radiative‐path‐aware retrieval framework: it first defines three cloud scenarios (cloudless, partial and overcast) based on distinct radiative transfer paths and then identifies them using satellite texture features and a hierarchical support vector machine (SVM). Then, for each scenarios, employs scenario‐specific graph networks to model spatial contexts, with joint optimisation of the observation domain and model parameters. The retrieved high‐fidelity, wide‐area irradiance fields dynamically construct the graph for a forecaster consisting of a graph attention network (GAT) module and long short‐term memory (LSTM) module. This integrated pipeline reduces the mean absolute error (MAE) and root mean square error (RMSE) by approximately 13.99% and 11.79%, respectively, compared with a standard LSTM baseline. Moreover, it consistently outperforms representative state‐of‐the‐art spatiotemporal graph models, confirming that the retrieved wide‐area surface solar irradiance (SSI) provides effective and reliable spatiotemporal inputs for correlation‐aware ultra‐short‐term PV power forecasting.