An efficient approach for regional photovoltaic power forecasting optimization based on texture features from satellite images and transfer learning

Accurate and efficient forecasting of regional photovoltaic (PV) power is essential for enhancing the stability of PV electricity supply and increasing its market share. Recent advancements have integrated features from satellite and ground observations, and hybrid neural network-based models have demonstrated impressive performance. However, challenges remain: spatial features extracted directly from satellite images often lack detail, and the majority of existing forecasting methods require extensive power data samples. Consequently, forecasting accuracy suffers from phase lags, particularly under conditions of high cloud cover change rates, and computational burdens are exacerbated by the vast number and dispersed nature of regional PV installations. To address these issues, this study proposes an innovative spatial–temporal feature that combines reconstructed texture features (TFs) from satellite images with ground observations to reduce forecasting phase lags and enhance accuracy. Additionally, a forecasting module that integrates 3D CNN, ConvLSTM, and ResNet within a transfer learning strategy effectively mitigates computational burdens. The obtained results indicate that forecasting accuracy has improved by up to 72% in terms of RMSE and 26% in terms of lag ratio compared to normal features, while computation time has been reduced tenfold compared to the traditional forecasting strategy. The robustness of the proposed algorithm has also been validated with various proportions of training data, ensuring reliability under diverse operational conditions.