基于BERT的电网现场作业风险自动评级方法

Automatic Risk Rating Method for Power Grid Field Operation Based on BERT

电网现场作业风险等级影响着作业的安全性和经济性,目前风险等级评级主要依赖于人工评级,存在一定的错误率。以国网公司的作业风险分级表与实际现场历史文本库作为研究对象,在充分分析各自文本特点的基础上,提出了基于BERT(bidirectional encoder representations from transformers)模型的电网现场作业风险自动评级方法。考虑作业风险分级表的概括性,提出对作业风险分级表进行文本增强;考虑实际现场历史文本库中存在一定比例的冗余内容和错误评级,提出冗余文本去除方法和基于文本挖掘技术的风险等级纠错方法;以经过上述方法处理后的2类现场作业文本为样本,构建了基于BERT单文本主题分类微调模型的文本分类模型,实现了对电网现场作业风险的自动评级。算例对比了该文模型与其他分类模型在不同样本集下的评级效果,表明BERT模型能够更全面地获取作业文本的语义特征,分级表的文本增强方法和历史库的风险等级纠错方法能提升评级效果,而冗余文本去除会导致作业文本样本数量减少较多,反而降低评级效果。上述结果表明,该文方法能够实现电网现场作业文本的自动评级,并具有较高的准确性。

The risk grading of the power grid field operation affects the safety and economy of operations.The current risk grading is mostly given manually,which contains a certain proportion of wrong ratings.Taking the operation risk classification table proposed by the State Grid Corporation and the actual field operation history text data set as the research objects,the automatic risk rating method of the power grid field operation based on the BERT model is proposed on the basis of the analysis of the respective text characteristics.Considering the generality of the operation risk classification table,we propose a text enhancement method for the operation risk classification table;taking in to account the existence of a certain proportion of redundant texts and the wrong ratings in the actual field operation history text database,we propose a redundant text removal method and an correction method of the risk rating based on text mining technology.Taking the above two types of texts processed by the above methods as the sample set,we construct a risk rating model based on the BERT single-text topic classification fine-tuning model to predict the risk grade,which realizes the automatic risk rating of the field operations.Comparing the rating effect of the BERT model with other classification models under different sample sets,the results show that the BERT model is able to obtain the semantic features more comprehensively,that the text enhancement method of the operation risk classification table and the risk rating correction method of the history text database are able to improve the rating effect.However,the removal of the redundant text reduces the rating effect because of its excessive sample size reduction.The results show that the method proposed in the paper realizes the automatic risk rating for the power grid field operation texts and has a higher accuracy.