Wind power generation exhibits inherent stochasticity, posing challenges to grid safety and stability when integrated at large scale. Reliable wind power prediction mitigates these grid instability risks. The more accurate the predictions, the better for grid dispatch and economic efficiency. In this study, a CNN-LSTM-Attention model is innovatively applied to ultra-short-term wind power forecasting. Firstly, the wind power output characteristics are analysed using the Pearson correlation coefficient to quantify power related eigenvalues and spatiotemporal correlation scales. Secondly, refine the input factors of LSTM model by using the advantages of CNN feature extraction, and study the optimization-seeking and weighting method of the output parameters of LSTM by using Attention attention mechanism, and construct CNN-LSTM-Attention model. Finally, the findings of the present case study, which drew upon data from a wind farm, demonstrated that the combined model yielded lower prediction errors in comparison with CNN and LSTM alone. This finding suggests that the combination of these models enhances the correctness and trustworthiness of prediction. This model compensates for the shortcomings of traditional artificial neural networks.