A Convolutional Autoencoder-Based Approach for Feature Extraction and Label Optimization of Typical Loads in Distribution Networks

Extracting the latent value embedded in electricity load data constitutes one of the key challenges in the power industry. To address the difficulty faced by conventional clustering approaches in capturing the intrinsic features of high-dimensional load data, this paper proposes an optimized clustering method based on a one-dimensional convolutional autoencoder (1D-CAE). First, a 1D-CAE is employed to extract temporal features from daily customer load profiles through nonlinear dimensionality reduction, with the objective of minimizing reconstruction loss. Second, we introduce an improved Cayley orthogonal constraint to enhance the structural information of the clustering space, thereby optimizing the mapping of latent features and improving clustering stability. Third, a generative adversarial network (GAN) is integrated with K-means clustering to refine the cluster centers and fine-tune the encoder. Finally, the effectiveness of the proposed method is evaluated on real-world load datasets using three widely accepted internal validation metrics: the Davies–Bouldin index (DBI), the Calinski–Harabasz index (CHI), and the silhouette coefficient (SC). Experimental results demonstrate that the proposed approach significantly enhances both inter-cluster separability and intra-cluster compactness. The study confirms that the method can effectively identify and extract morphological characteristics of diverse load profiles, offering robust support for demand response and optimal dispatch in virtual power plants.