To improve the prediction accuracy of photovoltaic power， an EGO-CEEMDAN-VMD-BiGRU short-term photovoltaic power prediction model is proposed based on the dual data decomposition method of parameter optimization. Initially， the eel and grouper optimizer （EGO） algorithm is employed to determine the optimal parameters for complete ensemble empirical mode decomposition with adaptive noise （CEEMDAN）， thereby performing an initial decomposition of the photovoltaic dataset. Subsequently， the K-means clustering algorithm is utilized to categorize the modal components into high-frequency， medium-frequency， and low-frequency groups， effectively reducing redundancy among the components. Then， the EGO algorithm is used to optimize the parameters of variational mode decomposition （VMD）. Following this， the high-frequency component is decomposed for the second time to mitigate the non-stationarity of the sequence. Finally， bidirectional gated recurrent unit （BiGRU） is applied to predict the components derived from the two-stage decomposition process， with the final prediction result obtained through summation. Based on the dataset from a photovoltaic power plant in Ningxia， the EGO-CEEMDAN-VMD-BiGRU model was compared with the BiGRU， VMD-BiGRU， and CEEMDAN-VMD-BiGRU models. Under three distinct weather conditions， the average MAE was reduced by 68.93%， 55.84%， and 44.56%， respectively， while the RMSE was decreased by 68.23%， 53.38%， and 41.03%， respectively. The experimental results demonstrate that the proposed photovoltaic power prediction model exhibits high accuracy and stability， thereby holding practical significance for ensuring the safe and reliable operation of power systems.