Accurate prediction of photovoltaic (PV) power is of paramount importance in power system operation and scheduling. This study presents a method for predicting PV power based on a combination of global and local features. Initially, Conv2D and Conv1D convolutional layers are employed to extract both global and local features for PV power prediction. Subsequently, these extracted features are fed into an attention mechanism, which selects feature vectors with strong correlations for input into the Gated Recurrent Unit (GRU). Simultaneously, autoregressive prediction is performed for photovoltaic power generation. Ultimately, the obtained results are aggregated to yield the predicted outcomes. Finally, we validate the proposed method using a real dataset and compare its performance against other baseline models. The results indicate that the method proposed in this study yields a higher level of precision for predicting photovoltaic power generation.

solar energy,deep learning,Attention mechanism; Target tracking; Sparse representation; Salient model,Forecast