With the rapid development of the Electric Power Internet of Things (EPIoT), massive collection terminals such as distributed photovoltaic systems and smart meters generate complex heterogeneous and multi-attribute mixed operational data. Traditional clustering methods face challenges in handling such data, including insufficient adaptation to mixed attributes, inadequate characterization of security-sensitive features, and suboptimal grouping stability. To address these issues, this paper proposes a weighted mixed-attribute deep embedding clustering method for intelligent collection devices. Firstly, stacked denoising autoencoders are employed to achieve robust dimensionality reduction and deep feature extraction from high-dimensional mixed-attribute data. Secondly, a dynamic weighting mechanism based on information entropy is designed to quantify the importance of numerical and categorical attributes in secure grouping, and a weighted mixed-distance metric is constructed. Furthermore, a two-stage alternating optimization strategy is proposed: the first stage adopts t-distribution to calculate soft assignments, preserving the uncertainty of device membership; the second stage constructs a high-confidence target distribution to jointly update autoencoder parameters and cluster centers, achieving co-optimization of feature representation and grouping structure. Finally, the optimal number of clusters is adaptively determined using dual indicators of Sum of Squared Errors (SSE) and Silhouette Coefficient to ensure stable and reliable grouping results. Simulation experiments demonstrate that the proposed method outperforms existing comparison schemes in grouping accuracy, attack isolation rate, key update latency, and authentication overhead, providing effective support for collaborative control and intelligent decision-making in electric power Internet of Things. • Two-stage clustering for secure device grouping in Power IoT with hybrid-attribute data. • Dynamic weighting and alternating optimization for feature-clustering co-evolution. • Outperforms baselines on 750 terminals in accuracy, security and latency.
Sun et al. (Wed,) studied this question.