A large number of intermittent wind power grid will produce many uncertain problems, which will threaten the safe and stable operation of the power system. The scenario generation method describes the uncertainty problem as a deterministic problem, which is the key technology to solve the uncertainty problem of wind power. In this paper, an improved spatiotemporal power day scenario generation method based on Gaussian mixture clustering is proposed for conditional variational self-coding multi-wind farm. The hyperspherical distribution and the regularization of the maximum and minimum pinch Angle used in this method are helpful to obtain lower dimensional and more independent hidden variables and improve the accuracy of the model. In addition, the Gaussian mixture clustering technique used in this method can distinguish historical wind conditions and generate daily scenes of specified wind conditions. This makes the generation model more flexible and the generated scenarios more diverse.