Abstract: In order to respond to the national goal of "double carbon", a combined prediction model based on CNN-BiLSTM with dual-stage attention mechanism for preliminary prediction and LightGBM for error correction is proposed to address the problem of wind power prediction errors affecting the safe and stable operation of power grids. The model first uses a convolutional neural network(CNN) combined with an attention mechanism to form a feature attention module to adaptively extract important features of wind power, then uses a BiLSTM network combined with an attention mechanism to form a temporal attention module to make preliminary predictions of wind power, and finally uses LightGBM to construct an error correction model to correct the preliminary prediction results. Using the mean absolute error(MAE), root mean square error(RMSE) and coefficient of determination(R²) as experimental evaluation metrics, the results show that the combined model predicts significantly better than BiLSTM and CNN-BiLSTM models.