To reduce the impact of wind power fluctuations on power systems, a short-term power forecasting method for large-scale wind farm clusters is proposed, which accounts for spatiotemporal correlation and simultaneously outputs short-term power predictions for all wind farms. First, an evaluation index that comprehensively considers the spatial correlation of wind speed and direction is proposed, and a graph topology structure is further established to characterize the spatiotemporal correlation of wind farm clusters. Then, a deep residual graph attention network is constructed to mine the spatiotemporal correlation features between multiple wind farms, preserving the valuable spatiotemporal information embedded in the data during training. Finally, a false prediction evaluation index is proposed to assess the false prediction components of the predicted power at the station when aggregated into cluster prediction power, making a fairer evaluation of cluster prediction results. Experiments are conducted using a cluster composed of 20 wind farms in Jilin province, China. Results show that the proposed wind power forecasting model achieves a day-ahead power prediction accuracy rate of 91.68%.