In a certain environment, the regional wind farms distribute irregularly. the traditional CNN prediction method can not reflect the distribution state and influence relationship of regional wind farms, and it is difficult to accurately predict wind speed. To solve this problem, GCN is used for feature modeling, and the connected graph and weight matrix are established according to the topology of multiple wind fields and the cross correlation coefficient of wind speed in each region. Secondly, Depending on the time dynamic characteristics of wind speed, the improved parallel convolution structure is used to obtain the correlation of wind speed series in multiple time periods under the same wind field. Thirdly, Using the spatial correlation and delay effect of wind speed, the spatio-temporal characteristics of wind speed in different regions are aggregated by second-order aggregation method. Finally, the verification of a regional wind field data shows that this method can extract spatio-temporal characteristics and improve the prediction performance in multi wind field ultra short-term wind speed prediction at the 0-4h prediction scale.