To address the misclassification issues in existing density peak clustering algorithms caused by significant local fluctuations in data or uneven distribution of load categories, a power load curve clustering method based on singular spectrum analysis and improved density peak clustering algorithm is proposed. Firstly, singular spectrum analysis is employed to decompose the original power load data into a low-frequency component containing the main contour information and a high-frequency component representing noise. Secondly, the local density of the density peak clustering algorithm is redefined by integrating the ideas of K-nearest neighbors and natural nearest neighbors, followed by clustering the low-frequency component. Finally, applied to real-world power load datasets, the proposed method is compared with other clustering algorithms, and case study results validate its effectiveness in real data.