To address the challenges of existing deep learning prediction models, such as long training times and a tendency to fall into local optimum, this paper proposes an innovative KOA-driven VMD-CNN-BiGRU-Attention method for short-term photovoltaic （PV） output power prediction. Firstly, the Pearson correlation coefficient is employed to identify key meteorological factors. Then, the grey relational analysis （GRA） method is used to determine the historical similarity days for the predicted days. The Keplerian Optimization Algorithm （KOA） is then used to optimize the parameters of variational mode decomposition （VMD）, which decomposes the output sequences of historical similarity days to generate a high-quality training sample set. Finally, the VMD-CNN-BiGRU-Attention model, driven by KOA, is constructed to achieve accurate PV output power prediction. Practical applications at PV power stations in Yunnan and Gansu show that the model achieves RMSE values of 0.2540 MW and 2.7981 MW, and MAPE values of 0.0234 and 1.1699, respectively. Compared with other combined prediction models, the proposed KOA-driven VMD-CNN-BiGRU-Attention model demonstrates superior ability to capture spatiotemporal features, offering significant improvements in prediction accuracy and stability. These results highlight the broad application potential of the model in PV power generation prediction.

PV output prediction  /  Keplerian optimization algorithm （KOA）  /  variational mode decomposition （VMD）  /  Attention mechanism  /  CNN-BiGRU-Attention combined model