Abstract: Mining the multivariate correlation of wind farms has positive effect on improving the prediction accuracy in the medium and long term. In view of the shortcomings of Transformer model in capturing multi-variable correlation,a multi-variable medium and long term prediction model considering multiple correlation is proposed. Firstly,multivariate independent embedding（MIE）is used to model multiple variables of the wind farm. Then,two-dimensional probabilistic sparse attention（TPSA）is used to extract the feature information between time and variables. Finally,the multi-scale feature information is aggregated by multilayer-style encoder-decoder（MED）to output the prediction results at one time. Example analysis shows that compared with LSTM model,Transformer model and Informer model,the mean square error of the proposed model decreases by 42.58%～66.83%,32.58%～53.49% and 14.38%～30.92%respectively in each prediction time. The effectiveness of the proposed improvement is verified and analyzed by ablation experiments.