Aiming at the scare historical data of newly-built wind farms and difficulty in accurately predicting the power, a wind power forecasting method is put forward with few data considering the spatial coupling relationship based on the historical data of neighboring wind farms. First, spatial attention is proposed to quantify the influence degree of adjacent wind farms on the target wind farms, and the corresponding wind farms are weighted. Then, the hidden temporal correlation information in wind power sequence is obtained by the gated recurrent unit networks. Finally, the weight and offset of output layers are optimized by the crisscross optimization algorithm. The experiment results from the actual data of four neighboring wind farms in Inner Mongolia show that the proposed method can effectively quantify the impact of wind farms and improve the prediction accuracy of wind farms with fewer data.