Textual Embedding Optimised CCLLM Model for Work Order Text Classification in Smart Grid

Abstract

Text classification is a core task in natural language processing with significant implications across industries. In the context of grid work order classification within the power sector, it directly impacts the monitoring of power equipment status, the efficiency of fault diagnosis and the stability of power supply services. However, issues such as spelling errors in grid work orders pose significant challenges to traditional classification methods. While deep learning models like BERT have advanced semantic understanding, they still face limitations when handling texts with spelling errors. To address this issue, we propose the CCLLM model, which optimises ChineseBERT by integrating semantic feature prompting design and intelligent knowledge mining techniques. This enhancement improves the accuracy and robustness of grid work order classification. Experimental results demonstrate that CCLLM achieves notable improvements in both accuracy and robustness compared to models like BERT and ERNIE, and these findings are validated through ablation studies.