Enhanced Topic and Hotspot Analysis for Power Work Orders Using SBERT-LDA and Improved K-Means
Jinxin Si, Xinping Wu, Fuyong Sun, Xinzhou Geng, Qiuhe Ma, Xiuhuan Zang, Wenjing Zeng, Chunhui Zhao, Yu Miao Gao, Junxi Cao · 2024
In the electric power industry, the efficient processing and analysis of power work order data has emerged as a significant and urgent challenge. This paper addresses the challenges of processing and analyzing short texts in power company customer inquiries, particularly in the context of power work orders. These inquiries often contain sparse and domain-specific terminology, making traditional topic identification methods, such as Latent Dirichlet Allocation (LDA), ineffective. To overcome these limitations, a hybrid model is proposed, which combines Sentence-BERT for semantic representation and LDA for topic modeling. Furthermore, we enhance hotspot analysis using a Canopy-based improved K-means clustering algorithm, which optimizes clustering stability and accuracy. The performance of the proposed model is evaluated using the 95598 power work order dataset, and it exhibits superior topic coherence and clustering quality in comparison to conventional methods.