融合自注意力和时空卷积网络的空气质量预测方法  

Air quality prediction methods based on self-attention and spatiotemporal convolutional network

作　　者：杨国亮 余华声 夏金龙 黄经纬 YANG Guoliang;YU Huasheng;XIA Jinlong;HUANG Jingwei(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341000,Jiangxi,China)

机构地区：[1]江西理工大学电气工程与自动化学院,江西赣州341000

出　　处：《安全与环境学报》2023年第12期4580-4589,共10页Journal of Safety and Environment

基　　金：江西省教育厅科技计划项目(GJJ190450);江西省教育厅科技项目(GJJ180484)。

摘　　要：准确预测空气质量对人体健康和社会环境治理具有积极影响。选取北京市35个监测站点自2019年2月1日至2020年1月31日的逐小时空气质量监测数据与气象数据,利用最大信息系数(Maximum Information Coefficient,MIC)对多个监测站点进行空间相关性分析,并通过升降维的卷积方式进行特征提取与信息整合,得到具有时空相关性的输入特征信息;构建融合自注意力(Self-Attention,SA)的时空卷积网络模型(Spatiotemporal Convolutional Network-Self-Attention,STCN-SA),对选定的中心站点进行未来1 h的PM_(2.5)质量浓度预测。结果表明:与常见的循环神经网络(Recurrent Neural Network,RNN)和长短期记忆网络(Long Short-Term Memory Network,LSTM)模型相比,STCN-SA网络模型在平均绝对误差(E_(MAE))、均方根误差(E_(RMSE))和决定系数(R~2)表现出更好的预测性能。此外,该预测模型适用于不同空间位置的监测站点,具有良好的可移植性,可为预测空气污染物质量浓度提供重要参考。PM_(2.5) mass concentration is a key indicator for measuring the degree of air pollution,therefore,accurately predicting the concentration of air quality PM_(2.5) has a positive impact on human health and the environmental governance of society.In this study,a statistical modeling method was proposed for predicting hourly PM_(2.5) mass concentration.It can capture historical spatiotemporal data efficiently and without causing problems of information leakage from the future to the past and has an excellent capability of the non-linear fitting.Firstly,selecting hourly air quality data and meteorological data of 35 monitoring stations in Beijing from February 1,2019,to January 31,2020,as the dataset.For missing data and incomplete data,the data is pre-processed using interpolation methods with upper and lower data padding.Secondly,using maximum information coefficient(MIC) is to analyze the spatial correlation of multiple monitoring stations.Feature extraction and information integration through the convolution method of dimensional upgrading and downscaling are carried out to obtain the input feature information with spatiotemporal correlation.Furthermore,a spatiotemporal convolutional network model combining self-attention(STCN-SA) was constructed to predict the future hourly air quality PM_(2.5) mass concentration for selected central stations.In addition,a testing dataset was used to predict and evaluate the performance of the proposed model and to compare it with existing models.Compared with common recurrent neural network(RNN) and long short-term memory network(LSTM) models that merely consider temporal correlation,the results show that the proposed model performs better prediction performance in terms of mean absolute error(E_(MAE)),root mean square error(E_(RMSE)) and coefficient of determination(R~2).In summary,the air quality prediction method proposed in this study integrates temporal and spatial correlation,which not only can effectively capture the trend of air pollutant concentration changes in advanc

关 键 词：环境工程学 空气质量预测 PM_(2.5)质量浓度 时空相关性 自注意力 时间卷积网络 

分 类 号：X513[环境科学与工程—环境工程]