Abstract: Mining the causal correlation between wind power and wind energy, as well as the spatio-temporal correlation among multiple wind farms from multi-source, multi-dimensional, and multi-modal wind power data, is an effective way to improve the accuracy of wind power prediction. A data-driven hybrid deep-learning model is proposed. First, in the data preprocessing stage,the scene of the wind farm operating at full power output and wind speed overflowing is regarded as the abnormal and fault state of wind power prediction, and the corresponding data-cleaning method is proposed to enhance the correlation between the wind speed and wind power. Then, aiming at the causal correlation between meteorological data and wind power, multi-channel convolution is designed to mine the coupling relationship; aiming at the spatio-temporal correlation between adjacent wind farms,the multi-scale convolution is designed to mine wind power characteristics with different spatio-temporal scales, and these spatio-temporal characteristics are weighted, fused, and reduced through the 1× 1 convolution layer. Finally, the fusion characteristics are input into the gated recurrent network to mine the time correlation and output the prediction results of wind power for multiple wind farms. The effectiveness of the proposed model is verified by real datasets.