基于融合注意力机制的电力工程数据特征提取算法设计  

Design of power engineering data feature extraction algorithm based on fusion attention mechanism

作　　者：张可抒 吴丹 沈蕾 ZHANG Keshu;WU Dan;SHEN Lei(Changzhou Changgong Electric Power Design Institute Co.,Ltd.,Changzhou 213000,China)

机构地区：[1]常州常供电力设计院有限公司,江苏常州213000

出　　处：《电子设计工程》2024年第17期141-145,共5页Electronic Design Engineering

基　　金：常州科技计划项目(X2022CZCG2B0643)。

摘　　要：针对传统电力工程造价预测主要依靠经验进行估计所存在的不足,文中基于注意力机制和改进的神经网络提出了一种电力工程造价数据特征提取与预测算法。该算法以历史造价数据为输入,利用DBN网络进行全局特征提取,并使用模拟退火算法对网络神经元数量及参数加以优化。对于DBN网络特征提取能力不足的问题,采用Bi-GRU算法实现对数据局部特征的提取,且利用多头注意力机制对特征进行权重分配,以提高算法的预测性能。在实验测试中,所提算法的MAPE、RMSE、MAE误差指标相较于XGBoost算法分别降低了0.0086、56和31,工程应用测试中的造价预测误差则在4.02%以内,证明该算法的综合性能良好,达到了设计目的。Aiming at the defect that traditional electric power project cost prediction mainly depends on experience,this paper proposes a feature extraction and prediction algorithm of electric power project cost data based on attention mechanism and improved neural network.The input data of the algorithm is historical cost data,and the DBN network performs global feature extraction.Simulated annealing algorithm is used to optimize the number of neurons and parameters of the network.For the lack of feature extraction ability of the DBN network,Bi⁃GRU algorithm is used to extract local features of the data,and multi head attention mechanism is used to assign weight to the features,so as to improve the prediction performance of the algorithm.In the experimental test,the error indexes of MAPE,RMSE and MAE of the proposed algorithm are 0.0086,56 and 31 lower than that of XGBoost algorithm,and the cost prediction error in the engineering application test is within 4.02%,which proves that the comprehensive performance of the algorithm is good and reaches the design purpose.

关 键 词：注意力机制 模拟退火算法 深度置信网络 双向GRU网络 数据预测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TN929.5[自动化与计算机技术—计算机科学与技术]