基于改进DeBERTa-DPCNN神经网络的电力工单识别技术研究OA
Research on Power Grid Work Order Recognition Technique Based on Improved DeBERTa-DPCNN Neural Network
针对当前电力工单文本识别技术的术语覆盖不足、短文本特征提取不充分及鲁棒性低等问题,提出一种融合改进解耦注意力的解码增强中文双向编码器表征模型(decoding-enhanced bidirectional encoder representations from transformers with disentangled attention-Chinese,DeBERTa-Chinese)与深度金字塔卷积神经网络(deep pyramid convolutional neural network,DPCNN)的电力工单智能识别模型(DeBERTa-DPCNN for power work orders recognition,DeDPC-P).DeDPC-P基于中文预训练的DeBERTa-Chinese模型,使用电力领域术语扩充词典并引入对抗训练策略进行微调,增强模型对电力术语的语义理解与抗噪声干扰能力;同时结合DPCNN多层级卷积与残差连接结构,充分提取工单文本的深层语义特征,构建端到端的工单分类模型.实验结果表明,在包含27 680条真实电力工单数据集中,DeDPC-P的宏F1值达0.941 0,较单一DeBERTa模型和DPCNN模型分别提升3.14%和6.09%.DeDPC-P的宏精确率达0.941 5,且在含噪声工单场景下表现出较高的鲁棒性.
Aiming to address the challenges of inadequate term coverage,insufficient feature extraction from short texts,and low robustness in the current electric power work order recognition technology,we propose an intelligent recognition model for power grid work orders,named DeDPC-P.DeDPC-P integrates an enhanced version of the DeBERTa-Chinese model with a Deep Pyramid Convolutional Neural Network(DPCNN).This method leverages a pre-trained Chinese DeBERTa model,enhanced by expanding its lexicon with domain-specific power terminology and fine-tuning it through adversarial training strategies.This approach strengthens the model's semantic understanding of power-related terms and its robustness against noise interference.Furthermore,the DPCNN employs multi-layer convolutional structures and residual connections to comprehensively extract local key features and long-distance dependencies from work order texts,thereby constructing an end-to-end classification framework.Experimental results demonstrate that,on a real-world dataset of 27 680 power work orders,DeDPC-P achieves a macro-F1 score of 0.941 0,improving by 3.14%and 6.09%compared to standalone DeBERTa and DPCNN models,respectively.DeDPC-P achieves a macro precision of 0.941 5 and exhibits strong robustness in noisy work order scenarios.
刘嘉熹;高丽媛;付昀夕;郭光来;张沛尧
国网信息通信产业集团有限公司,北京 102211国网信息通信产业集团有限公司,北京 102211国网信息通信产业集团有限公司,北京 102211国网信息通信产业集团有限公司,北京 102211国网信息通信产业集团有限公司,北京 102211
信息技术与安全科学
工单识别DeBERTa-Chinese模型深度金字塔卷积神经网络模型电力系统
power grid work orderDeBERTa-ChineseDPCNNpower system
《山东电力技术》 2026 (8)
111-124,14
国网信息通信产业集团有限公司科技创新项目"面向新型电力系统的数字化配电网精准规划平台、协同作业枢纽与能源绿证交易研究专项"(SGIT0000KJJS2400504). Science and Technology Project of State Grid Information&Telecommunication Group Company"Research Project of Digitalized Distribution Grid Precision Planning Platform,Cooperative Operation Hub and Energy Green Certificate Transaction for New Electricity Systems"(SGIT0000KJJS2400504).
评论