中国电力 ›› 2025, Vol. 58 ›› Issue (11): 156-163.DOI: 10.11930/j.issn.1004-9649.202411021

• 新型电网 • 上一篇    下一篇

基于双层注意力机制的电力文本分类模型

武同心1(), 纪鑫1,2(), 杨成月1(), 陈屹婷1, 杨智伟1   

  1. 1. 国家电网有限公司大数据中心,北京 100031
    2. 北京航空航天大学,北京 100191
  • 收稿日期:2024-11-06 修回日期:2025-06-20 发布日期:2025-12-01 出版日期:2025-11-28
  • 作者简介:
    武同心(1986),男,硕士,高级工程师,从事电力系统知识图谱构建研究,E-mail: tongxin-wu@sgcc.com.cn
    纪鑫(1984),男,硕士,高级工程师,从事电力大数据、云计算、人工智能研究与开发,E-mail:xin-ji@sgcc.com.cn
    杨成月(1977),男,博士,高级工程师,从事摄影测量与遥感等研究,E-mail:chengyue-yang@sgcc.com.cn
  • 基金资助:
    国家电网有限公司科技项目(基于图神经网络和图深度学习的电力知识抽取技术研究,52999021N005)。

A Power Text Classification Model Based on Dual-layer Attention Mechanism

WU Tongxin1(), JI Xin1,2(), YANG Chengyue1(), CHEN Yiting1, YANG Zhiwei1   

  1. 1. State Grid Corporation of China Big Data Center, Beijing 100031, China
    2. Beijing University of Aeronautics and Astronautics, Beijing 100191, China
  • Received:2024-11-06 Revised:2025-06-20 Online:2025-12-01 Published:2025-11-28
  • Supported by:
    This work is supported by Science and Technology Project of SGCC (Research on Power Knowledge Extraction Technology Based on Graph Neural Network and Graph Deep Learning, No.52999021N005).

摘要:

电力领域中存在大量的中文文本数据,传统文本挖掘方法存在分词难度大、文本特征提取具有局限性、处理文本中的复杂关系时效果不好等问题,严重限制了对电力信息的深度理解与分类。基于此,提出了一种结合文本卷积网络(text convolutional neural networks,TextCNN)与注意力(Attention)机制的电力文本分类模型,对输入层、TextCNN层、第1个注意力层、池化层、第2个注意力层以及输出层进行了分层优化设计,并通过实验对模型进行了验证。结果显示,TextCNN-Attention模型在电力文本数据集上的文本分类准确率达到了96.8%,精确率达到了86.3%,召回率达到了90.3%,综合评价值达到了88.2%,表明TextCNN-Attention模型可有效提升电力文本分类的效率,为深度学习算法在电力文本分类领域的应用提供了经验借鉴。

关键词: 深度学习, 电力领域, 文本分类, 卷积神经网络, 注意力机制

Abstract:

There are a large amount of Chinese text data in the electric power field, traditional text mining methods are faced with problems such as difficulty in word segmentation, limitations in text feature representation, and poor performance in handling complex relationships in text, which limit the deep understanding and classification of power information. This paper proposes a power text classification model that combines text convolutional neural networks (TextCNN) and Attention mechanism. A hierarchical optimization design was carried out for the input layer, TextCNN layer, first attention layer, pooling layer, second attention layer, and output layer, with experimental validation conducted to verify the model's performance. The results show that the proposed TextCNN-Attention model achieved a text classification accuracy of 96.8%, with a precision of 86.3%, a recall of 90.3%, and a F1 score (comprehensive evaluation metric) of 88.2% on the power text dataset, demonstrating the superior performance of the TextCNN-Attention model in processing power texts. This study can provide valuable experiences for application of deep learning in power text classification.

Key words: deep learning, electricity sector, text classification, convolutional neural networks, attention mechanism


AI


AI小编
您好!我是《中国电力》AI小编,有什么可以帮您的吗?