Abstract: In a wind farm where the wind turbines are arranged irregularly, the distribution of many wind turbines in different spatial locations forms a point cloud instead of a regular rectangular grid. The point cloud is irregular and unordered, which can represent the geographical location distribution of multiple wind turbines in any wind farm. However, the point cloud data cannot constitute the input of the highly regular grid for the convolutional neural network（CNN）. It is difficult for the convolution operator to learn the spatial local correlation in the point cloud. If the irregular point cloud is directly mapped to a grid arrangement for regular convolution, the original spatial information of the point cloud will be lost. Therefore, the point CNN is used for spatial correlation extraction, then the simple recurrent unit is used to extract the temporal correlation information. And then the spatio-temporal correlation of point cloud data can be obtained. At the same time, the spatial features on multiple scales are extracted and integrated when designing the point CNN. Finally, the simulation results of multiple wind turbines with actual point cloud distribution show the effectiveness of the proposed forecasting model.