Sky image-informed hybrid model for ultra-short-term prediction of photovoltaic power output

Ensuring the accuracy of solar power forecasts is essential for maintaining stable electrical grids and implementing effective real-time economic dispatch strategies. This study introduces a pioneering approach to 15-minute ultra-short-term photovoltaic power forecasting, integrating sophisticated signal processing methods with advanced neural networks and incorporating data from sky imagery. A novel feature extraction technique from sky images has been developed, linking pixel colors to their physical meanings to enhance the interpretability of the extracted features. The research presents a hybrid forecasting model that combines Ensemble Empirical Mode Decomposition (EEMD) with the Gated Recurrent Unit (GRU), significantly improving the accuracy of the forecasts and including the sky image feature (SIF) resulting in a reduction of 16.5% in NRMSE and 4.9% in MAPE, and a notable enhancement of 2.7% in R2 when juxtaposed with the model that does not utilize SIF. Compared to other hybrid models, EEMD-GRU has a higher prediction accuracy with an average reduction of 28.82% and 36.97% in NRMSE and MAPE, respectively. These improvements underscore the method's effectiveness in enhancing the precision of photovoltaic power predictions by leveraging sky imagery and a deep learning model based on signal decomposition.