Photovoltaic Nowcasting With Bi-Level Spatio-Temporal Analysis Incorporating Sky Images

Today many large-scale photovoltaic (PV) plants have been equipped with sky imaging systems. The sky images contain abundant spatio-temporal information of the local climate condition at the PV plant. Meanwhile, power outputs of the PV systems located at different sites in a certain geographical area also exhibit spatio-temporal correlation. It is of interest to improve the accuracy of PV nowcasting by fully utilizing spatio-temporal information embedded in both sky images and PV output measurements of distributed PV systems. In this article, we incorporate the above two aspects into a unified framework and propose the Bi-level spatio-temporal (BILST) PV nowcasting model. The proposed model learns features from local spatio-temporal information embedded in sky images, global spatio-temporal correlations embedded in PV output datasets of a number of distributed PV systems and weather characteristics embedded in exogenous dataset simultaneously. Then the obtained three types of hidden features are aggregated and applied to predict the PV output at the PV site of interest. Experiments using real-world datasets show that the proposed BILST model can enable sky images to contribute to the PV nowcasting task and achieve the desirable accuracy of PV nowcasting consistently.