The intermittent, variable and stochastic nature of photovoltaic power generation poses a major challenge to grid integration. To address this problem, a combined prediction model CNN-LSTM-Attention (CLA) combining convolutional neural network (CNN), long short-term memory (LSTM) and Attention mechanism was proposed, which can effectively capture the potential patterns in time series. An improved multivariate universe optimization algorithm was used to optimize the LSTM parameters to enhance the predictive ability and robustness of the model. It was worth mentioning that the model design did not need to rely on complex data preprocessing steps, demonstrating good versatility and practicality. The introduced Attention mechanism empowered the model to adaptively focus on key historical information, thus significantly improving the prediction accuracy. The experimental results show that the model outperforms other models in terms of prediction accuracy and stability, with a coefficient of determination (R²) of 0.978, which exceeds the traditional models of 0.032 and 0.039, and when compared with the recently developed optimization models, it is 0.028 and 0.023 higher than the sparrow search algorithm-CNN-LSTM(SSA-CL) and dung beetle optimizer-CNN-LSTM(DBO-CL)models, respectively, and the error is significantly smaller than that of the other models. The application potential of this method in real PV power forecasting is verified, providing a strong technical support for future grid scheduling and operation.