基于高阶图卷积网络的城市空气质量推断模型  

A high-order graph convolutional network for urban air quality inference

作　　者：陈杰[1] 许镇义 Chen Jie;Xu Zhenyi(Department of Automation,University of Science and Technology,Hefei 230026,China;Institute of Artificial Intelligence,Hefei Comprehensive National Science Center,Hefei 230088,China)

机构地区：[1]中国科学技术大学自动化系,安徽合肥230026 [2]合肥综合性国家科学中心人工智能研究院,安徽合肥230088

出　　处：《信息技术与网络安全》2021年第4期33-41,45,共10页Information Technology and Network Security

基　　金：国家重点研发计划项目(2018AAA0100800,2018YFE0106800);国家自然科学基金(61725304,61673361);安徽省科技重大专项(912198698036);中央高校基本科研业务费专项资金(WK2100000013)。

摘　　要：能否精确地预测城市区域空气质量分布,对于政府环境治理以及人们日常预防等方面,具有重要的意义。该问题面临的挑战是:一是不同区域的空气质量分布具有时空交互性;二是空气质量分布受到外部因素的影响。通用化卷积神经网络以处理任意图结构数据,成为近些年来研究的热点之一,将城市空气质量预测问题可制定为时空图预测问题。基于提出的高阶图卷积网络,设计了一种有效的空气质量推断模型。该模型可以捕获空气质量分布的时空交互性和提取外部影响因素特征,从而精确预测空气质量分布。通过验证现实北京市空气质量数据,结果表明提出的模型远远优于目前已知的通用方法。Whether it can accurately predict the air quality distribution is of great significance to the government′s environmental governance and people′s daily health prevention.This problem is challenging for the following reasons:(1)The air quality distribution in different regions has temporal and spatial interaction;(2)The air quality distribution is affected by external factors.In recent years,generalized convolutional neural network(CNN)is one of the research hotspots to process arbitrary graph structured data,so the fine-grained air quality forecasting problem in urban areas is formulated as an urban spatio-temporal graph prediction problem.Based on the proposed high-order graph convolution,we design an effective air quality inference model for inferring the air quality distribution,which could capture the spatio-temporal interaction of air quality distribution and extract external influential factor features.Through the verification of Beijing air quality data,experimental results show that proposed approach far outperforms known baseline methods.

关 键 词：空气质量 时空交互性 图卷积网络 半监督学习 

分 类 号：P41[天文地球—大气科学及气象学]