摘要
细颗粒物(PM2.5)与大气环境和人类生活息息相关。城市中PM2.5监测站数量有限,无法提供细粒度PM2.5浓度,而大多数现有的PM2.5浓度推断方法缺乏根据动态时空特征建立多阶相关系数矩阵的能力。为此提出了一种基于注意力机制的PM2.5多阶图卷积网络推断模型(MOSTGCNInf)。该模型在利用图神经网络提取特征关系的同时,采用注意力机制动态构建多阶节点的注意力系数矩阵,并进行时空特征融合来提升PM2.5浓度推断效果。在公开数据集上进行了对比实验,使用准确率和F_(1)值作为评价指标,并通过消融实验验证了方法的有效性。实验结果表明,MOSTGCNInf对PM2.5浓度推断结果有提升作用。
Fine particulate matter(PM2.5)is closely related to the atmospheric environment and human life.The number of PM2.5 monitoring stations in the city is limited,unable to provide fine-grained PM2.5 concentration,and most existing PM2.5 concentration inference methods lack the ability to establish a multi-order correlation coefficient matrix based on dynamic spatial and temporal characteristics.This paper proposed an attention based PM2.5 multi-order spatio-temporal graph convolutional network inference model(MOSTGCNInf).This model used a graph neural network to extract feature relationships,adopted an attention mechanism to dynamically construct an attention coefficient matrix of the order node and performed spatio-temporal feature fusion to improve the PM2.5 concentration inference effect.It carried out comparative experiments on the public data set,used accuracy and F_(1) value as evaluation indicators,verifying the effectiveness of the method through ablation experiments.Experimental results show that MOSTGCNInf can improve the results of PM2.5 concentration estimation.
作者
彭一非
杨维
Peng Yifei;Yang Wei(School of Electronic&Information Engineering,Beijing Jiaotong University,Beijing 100044,China)
出处
《计算机应用研究》
CSCD
北大核心
2022年第5期1491-1495,共5页
Application Research of Computers
关键词
PM2.5
相关系数矩阵
多阶图卷积
时空特征融合
注意力机制
PM2.5
correlation coefficient matrix
graph convolution
spatial-temporal feature fusion
multi-order attention mechanism
