In order to address the high randomness and volatility of photovoltaic power, a single prediction model often struggles to accurately analyze the fluctuation patterns in historical data, resulting in low prediction accuracy. To overcome this issue, a combined model for short-term PV power prediction is proposed, which incorporates Convolutional Neural Network-Gated Recurrent Unit (CNN-GRU) and an improved Sparrow Search Algorithm (ISSA) for optimizing the Xtreme Gradient Boosting (XGBoost) model. First, remove the Outlier in the historical data and normalize them, and use Principal Component Analysis (PCA) to select features, so as to better identify the key factors affecting photovoltaic power. Then, the CNN network is used to extract the spatial features of the data, and the GRU network is used to extract the temporal features. In response to the difficulty in manually configuring parameters and high randomness of the XGBoost model, ISSA is used to optimize the hyperparameters of the model. Finally, the results predicted by the two methods are reduced using the reciprocal error method while updating the weights to obtain new predicted values to complete the prediction of photovoltaic power. The experimental results show that the proposed CNN-GRU-ISSA-XGBoost combination model has stronger adaptability and higher accuracy.