基于深度特征融合网络的电力工程数据对比算法  

Comparison algorithm of power engineering data based on depth feature fusion network

作　　者：何洁明 何劲熙 HE Jieming;HE Jinxi(Guangzhou Power Supply Bureau of Guangdong Power Grid Co.,Ltd.,Guangzhou 510000,China;Guangzhou Junli Consulting Service Co.,Ltd.,Guangzhou 510000,China)

机构地区：[1]广东电网有限责任公司广州供电局,广东广州510000 [2]广州隽力咨询服务有限公司,广东广州510000

出　　处：《电子设计工程》2025年第6期39-43,共5页Electronic Design Engineering

基　　金：广东省科技计划项目(202274GKJJH76F29D)。

摘　　要：针对电力工程造价数据集数量少、质量差且难以应用于智能化的招投标数据校核的问题,文中基于改进的迁移学习模型提出了一种电力工程数据对比算法。针对传统迁移学习算法无法训练多重特征数据集的缺陷,采用JMMD函数对联合分布的差异进行度量,从而提高了算法训练的准确性。使用迁移学习算法对传统GAN进行一致性改进,并采用Cycle-GAN对数据进行训练。为了提升算法运行的效率,通过孪生神经网络对不同的输入数据进行预训练,得到自适应参数指导模型的训练。在实验测试中,所提算法运行速度、数据核查准确度在所有对比算法中均为最优,同时加入迁移学习模型后,训练少量样本数据集的性能下降相较原算法更慢,验证了算法改进的有效性。In response to the problem of limited quantity and poor quality of power engineering cost datasets that are difficult to apply to intelligent bidding data verification,this paper proposes a power engineering data comparison algorithm based on an improved transfer learning model.To address the limitation of traditional transfer learning algorithms being unable to train multiple feature datasets,the JMMD function is used to measure the differences in joint distributions,which improves the accuracy of algorithm training.Use transfer learning algorithms to improve the consistency of traditional GAN,and train the data using Cycle-GAN.In order to improve the efficiency of algorithm operation,twin neural networks are used to pre train different input data and obtain adaptive parameter guidance for model training.In experimental testing,the algorithm’s running speed and data verification accuracy were the best among all comparative algorithms.At the same time,adding a transfer learning model resulted in a slower performance decline compared to the original algorithm in training a small sample dataset,indicating the effectiveness of the algorithm improvement.

关 键 词：深度特征融合 深度迁移学习 循环对抗神经网络 孪生神经网络 电力工程数据 数据核查 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TN918.4[自动化与计算机技术—计算机科学与技术]