As the demand for smart grid construction increases, advanced power applications based on edge-cloud collaboration continue to increase. Among them, there are many data-driven artificial intelligence calculations and analyses, all of which are calculated and analyzed based on electric power big data. However, for the massive electric power big data, it is impossible to obtain more internally related information only by observing the data from the surface. To a certain extent, it directly affects the upper-level advanced applications. To solve this problem, this paper studies and proposes a curve-mean clustering algorithm for load big data, which is the most widely used load data in smart grid. By analyzing the advanced measurement infrastructure, the matrix low-rank property of load big data and the calculation of singular value, the curve mean clustering of load big data is realized, and the optimal determination method of cluster number is expounded. Experiments are conducted based on actual resident user load data and compared with the classic mean shift clustering algorithm. By calculating the average distance within the cluster, the average distance between clusters and the DI index, it is verified that the proposed method clustering is more accurate and the selection of cluster number is optimal. The research plays a very good role in basic analysis for improving the big data analysis capability and data quality of smart grid.