Abstract: In order to improve the accuracy of offshore wind power prediction and the reliability of its prediction results, a self-attention model integrating learnable wavelet is proposed, which effectively improves the accuracy of offshore wind power prediction and has a certain ability of forecasting process analysis. First, the wavelet decomposition is integrated with the deep learning model, so that the model has the ability to extract features from different frequency domains. Second, a sparse self-attention prediction network is constructed to extract global information features effectively and improve the prediction performance of the model. Then, a time-series "sensitivity" quantitative analysis method is proposed to evaluate the importance of offshore wind power parameter input in multiple dimensions, and to analyze the prediction mechanism reasonably to a certain extent. Finally, relevant experiments are carried out based on actual operation data of wind field. The experimental results show that, compared with the model, the proposed model has higher prediction accuracy in the ultra-short term prediction task of offshore wind power.