Mining the potential value of power load data is one of the main topics in the electric power industry. Aiming at the problem that it is difficult to extract the potential features of high-dimensional power load data by direct clustering methods， a one-dimensional convolutional autoencoder-based load feature optimization clustering method is proposed to optimize the clustering of typical load features. Firstly， a one-dimensional convolutional autoencoder is used to extract the time-series features of the user's daily load profile with the objective of optimizing the reconstruction loss to achieve the nonlinear dimensionality reduction of the data. Second， a cayley orthogonal constraint method for improving the clustering structure information is proposed to optimize the mapping of the features in the potential space and improve the clustering stability. Then the GAN-Kmeans clustering method is used to optimize the clustering center to fine-tune the coding part， so as to improve the accuracy of feature classification and identification. Finally，based on the user load dataset， three internal indicators， DBI， CHI and SC， are used to evaluate the effectiveness of the proposed method， which shows that the proposed method can effectively identify and extract the morphological features of various load curves， and it can provide a reliable support for demand response and optimal scheduling of virtual power plants.