A solar-region luminance extraction based scenario-adaptive PV power minute-scale forecasting method

With the increasing penetration of photovoltaic (PV) systems, achieving accurate PV power minute-scale forecasting has become essential. Reliable minute-scale forecasting requires meteorological forecasting information with high spatial and temporal resolution. Ground-based cloud images can provide real-time, fine-grained observations of cloud above PV sites and thus serve as a promising meteorological source for forecasting. Based on above characteristic, this study proposes a solar-region luminance extraction based scenario-adaptive PV power minute-scale forecasting method, utilizing ground-based cloud images. Firstly, a multi-scale ConvGRU network is developed to forecast future ground-based cloud images. Secondly, a solar-region luminance extraction method is proposed to precisely locate the solar region in cloud images and extract the luminance features within solar region. Thirdly, a scenario-adaptive forecasting method is introduced, which selectively incorporates cloud-image features into forecasting based on the scenario condition of samples, thereby improving overall accuracy of PV power forecasting. Experiments conducted on Stanford University’s SKIPP’D dataset demonstrate that the proposed method achieves a 15-min MAE accuracy of 96.8%, outperforming the benchmark cloud-fraction-based method by 16.8%. • Multi-scale ConvGRU network forecasts ground-based cloud images accurately. • Solar-region luminance extraction exacts key local features form cloud images effectively. • Scenario-adaptive forecasting method achieve accurate minute-scale forecasting utilizing cloud image.