The typical scene generation method has problems such as failing to fully consider the coupling relationship between uncertain factorslike wind, light and load, traditional clustering methods perform poorly on high-dimensional data sets, and the extracted typical scenes cannot well reflect the characteristics of the original data. To solve these problems,combining with Bayesian formula and multivariate Gaussian distribution, this paperfirstly uses Gaussian process regression (GPR) to model the coupling relationship of various uncertain factors in the power systemand generate the simulation operation data on the basis of improving the kernel function. Secondly,using the time series segmented typical scene extraction method, the total scheduling interval is divided into several sub-intervals. The center points are clustered respectively. Sub-interval weighted typical scenes are obtained and connected by Cartesian productto generate typical scene set of the full scheduling interval. Then, a method based on land movement distance (EMD) is applied to evaluate the extraction effect of typical scenes.Finally, it is verified that the extracted typical scenes can better retain the probability distribution characteristics of the original basic scene setand fully reflect the coupling relationship between uncertain factors in the original data set. The resultsshow that the typical scenes extracted by the method can be better reflect the characteristics of the original data.