Optimizing regional PV power forecasting using satellite cloud images and transfer learning strategy

• Texture features extracted from satellite cloud images enhance PV power forecasting. • A transfer learning strategy is employed to improve computational efficiency for regional PV stations. • Forecast accuracy and computational efficiency have been improved under various weather conditions. • Good generalization has been demonstrated with varying volumes of training data. As renewable energy, particularly regional photovoltaic (PV), becomes more prevalent in the power grid, accurate forecasting of its output is paramount for its efficient operation. However, challenges persist, including the lack of reliable data, inappropriate data usage, and the computational burdens stemming from the vast number and dispersed nature of regional PV installations. To address these problems, a regional PV power forecasting based on transfer learning and satellite cloud imagery is proposed. Firstly, an algorithmic architecture composed of gray-level co-occurrence matrix and random forest is established for extracting texture features (TF) from satellite cloud images and selecting the TF with the highest correlation to irradiance. Furthermore, an attention mechanism based on long short-term memory is employed to reconstruct these significant TF. These reconstructed TF are then integrated into the training data for the forecasting model, aiming to enhance the correlation between the TF and the forecasting outcome. Finally, a transfer leaning structure combine convolutional neural network and informer is taken as the forecasting model, and the self-organizing map and maximum mean discrepancy algorithm is utilized to distinguish the source and target PV stations. Both the single PV located in UK and the regional PV station located in China are analysis to verify the effectiveness, and several benchmark forecasting methods have been compared, the forecasting approach in this research demonstrated superior performance. With this strategy, the maximum improvement in accuracy reached to 8.7 in term of RMSE, the volume of training data is reduced to 60%, and the computational time reduced to 54.4%.