PM_(2.5)probabilistic forecasting system based on graph generative network with graph U-nets architecture  

基于图U-nets架构的图生成网络PM2.5概率预测系统

作　　者：LI Yan-fei YANG Rui DUAN Zhu LIU Hui 李燕飞;杨睿;段铸;刘辉(School of Mechatronic Engineering,Hunan Agricultural University,Changsha 410128,China;Institute of Artificial Intelligence and Robotics(IAIR),Key Laboratory of Traffic Safety on Track of Ministry of Education,School of Traffic and Transportation Engineering,Central South University,Changsha 410075,China)

机构地区：[1]School of Mechatronic Engineering,Hunan Agricultural University,Changsha 410128,China [2]Institute of Artificial Intelligence and Robotics(IAIR),Key Laboratory of Traffic Safety on Track of Ministry of Education,School of Traffic and Transportation Engineering,Central South University,Changsha 410075,China

出　　处：《Journal of Central South University》2025年第1期304-318,共15页中南大学学报(英文版)

基　　金：Project(2020YFC2008605)supported by the National Key Research and Development Project of China;Project(52072412)supported by the National Natural Science Foundation of China;Project(2021JJ30359)supported by the Natural Science Foundation of Hunan Province,China。

摘　　要：Urban air pollution has brought great troubles to physical and mental health,economic development,environmental protection,and other aspects.Predicting the changes and trends of air pollution can provide a scientific basis for governance and prevention efforts.In this paper,we propose an interval prediction method that considers the spatio-temporal characteristic information of PM_(2.5)signals from multiple stations.K-nearest neighbor(KNN)algorithm interpolates the lost signals in the process of collection,transmission,and storage to ensure the continuity of data.Graph generative network(GGN)is used to process time-series meteorological data with complex structures.The graph U-Nets framework is introduced into the GGN model to enhance its controllability to the graph generation process,which is beneficial to improve the efficiency and robustness of the model.In addition,sparse Bayesian regression is incorporated to improve the dimensional disaster defect of traditional kernel density estimation(KDE)interval prediction.With the support of sparse strategy,sparse Bayesian regression kernel density estimation(SBR-KDE)is very efficient in processing high-dimensional large-scale data.The PM_(2.5)data of spring,summer,autumn,and winter from 34 air quality monitoring sites in Beijing verified the accuracy,generalization,and superiority of the proposed model in interval prediction.城市空气污染给人体身心健康、经济发展、环境保护等各方面带来了极大困扰,预测空气污染的变化趋势可为治理和防治工作提供科学依据。本文提出一种考虑多站点PM2.5信号时空特征信息的区间预测方法,利用K最近邻(KNN)算法对采集、传输和存储过程中丢失的信号进行插值,保证数据的连续性。利用图生成网络(GGN)处理结构复杂的时间序列气象数据,在GGN模型中引入图U-Nets框架,增强其对图生成过程的可控性,有利于提高模型的效率和鲁棒性。此外,结合稀疏贝叶斯回归,改进传统核密度估计(KDE)区间预测的维数灾难缺陷。在稀疏策略的支持下,稀疏贝叶斯回归核密度估计(SBR-KDE)在处理高维大规模数据时非常高效。采用北京34个空气质量监测点的春、夏、秋、冬季PM2.5数据验证了所提模型的准确性、泛化能力以及区间预测的优越性。

关 键 词：PM_(2.5)interval forecasting graph generative network graph U-Nets sparse Bayesian regression kernel density estimation spatial-temporal characteristics 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]