OCAE-based feature extraction and cluster analysis of high-energy-consuming plant loads

• Published in: Computers & electrical engineering Vol. 119; p. 109579
• Main Authors: Zhou, Mengran, Kong, Weile, Hu, Feng, Zhu, Ziwei, Wu, Changzhen, Wang, Ling, Zhang, Qiqi, Zhou, Guangyao
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.11.2024
• Subjects: 1D convolutional autoencoder; K-means; Load clustering; Load feature extraction; Optimal convolutional autoencoder; K-means; Load feature extraction; Load clustering; Optimal convolutional autoencoder; 1D convolutional autoencoder
• ISSN: 0045-7906
• DOI: 10.1016/j.compeleceng.2024.109579

•Proposes a DL model that has better load feature extraction than 1D-CAE.•Proposes a DL model joint k-means load clustering method with excellent results.•Proposes a meteorological cycle weighting method to better fill in missing data.•Clustering analysis of real collected data on high-energy-consuming plant. Currently, in power systems and energy management, the extraction and clustering of load features play a critical role. It can support the planning, scheduling and equipment design of power systems, holding significant theoretical importance and practical benefits. Addressing the complexity of high-dimensional daily load data poses challenges in effectively capturing the intricate features of load sequence data and achieving optimal clustering outcomes. This paper proposes an optimal convolutional autoencoder (OCAE) model designed for good feature extraction. First, a large number of test comparisons demonstrate that the OCAE effectively mines deep feature data from load sequences and exhibits strong data reconstruction capabilities. Second, the OCAE model is compared with other autoencoders to validate its superior feature extraction performance. Finally, the deep feature information extracted by the OCAE encoder is utilized in the k-means clustering algorithm for cluster analysis and is quantitatively evaluated against other methods using Davies–Bouldin index (DBI) and silhouette coefficient (SC). The experimental results confirm the superiority of OCAE joint k-means clustering over improved 1D-CAE joint k-means clustering, with a reduction of approximately 12.3 % in DBI and an improvement of approximately 11.1 % in SC. The outstanding clustering outcomes achieved through OCAE joint k-means clustering have significant scientific implications for power system dispatch planning.