Aiming at the defects of traditional K-means algorithm in the power load curves clustering, such as sensitive to initial values and given the number of clusters, a dynamic K-means algorithm based on quantile radius is applied to the clustering analysis of daily load curves, and distributed improvement is made to optimize the computational efficiency. This algorithm combines two ideas:local clustering and global clustering in distributed clustering, and the central point obtained by K-means operation when k is repeatedly set as the fixed value for many times in the hierarchical K-means. Multiple K-means operations are assigned to different subsites, and the k-value of each K-means operation is changed. Then from the geometrical characteristics of the clusters, the concept of quantile radius is introduced. Quantile radius defines that the quantile of the distance between the sample point and the cluster center point represents the radius of the cluster. At the main site, the distance between the center points of each cluster with the quantile radius of the cluster is compared to filter and merge to get new clusters, so that the number of clusters can be quickly identified and a good initial center and result of the cluster are given. Finally, the daily load data of 606 users in a certain area and in a certain month is taken as the research object, and the effectiveness of the algorithm in the cluster analysis of power load curves is verified. This work is supported by National Key Research and Development Program of China (No. 2017YFE0112600).