基于后向气团轨迹的大气污染特征时序混合模型研究  

Research on time series hybrid model of air pollution characteristics based on backward air mass trajectory

作　　者：许睿 刘相阳 文益民 沈世铭 李建 Xu Rui;Liu Xiangyang;Wen Yimin;Shen Shiming;Li Jian(School of Computer Science and Information Security,Guilin University of Electronic Technology,Guilin,541004,China;Satellite Navigation Positioning and Location Service National&.Local Joint Engineering Research Center,Guilin,541004,China)

机构地区：[1]桂林电子科技大学计算机与信息安全学院,桂林541004 [2]卫星导航定位与位置服务国家地方联合工程研究中心,桂林541004

出　　处：《南京大学学报（自然科学版）》2022年第6期1041-1049,共9页Journal of Nanjing University（Natural Science）

基　　金：广西自然科学基金(2021GXNSFAA220056);广西重点研发计划(AB21196063);国家自然科学基金(62266014);桂林市重大成果转化基金(20192013-1);桂林电子科技大学大学生创新创业训练计划(202010595031)。

摘　　要：由于大气污染监测数据量大且具备长时序特征,深度学习领域已将其作为一种标准数据集使用.针对现有大气污染预测方法未能结合大气传输的物理机理和有效考虑污染物传输的时空特征等问题,提出一种基于多模型混合的特征与时序污染物预测模型,利用HYSPLIT(Hybrid Single-Particle Lagrangian Integrated Trajectory)计算后向气团运动轨迹,引入VGG(Visual Geometry Group)模型提取气团轨迹线的变化特征;将其与污染物和气象时序数据结合,输入LSTM(Long Short-Term Memory)模型,预测研究区域内的目标污染物浓度,以评估研究区域的空气质量状况.以桂林市61个空气质量监测站点的污染物和相关气象在线监测数据为基础,对模型性能进行了评估,将实验结果与几种先进的方法进行了比较.结果表明,提出的H-VGG-LSTM(HYSPLIT-VGG-LSTM)模型有效提高了大气污染物的预测准确度,其预测结果的RMSE(Root Mean Squared Error),MAE(Mean Absolute Error)和SMAPE(Symmetric Mean Absolute Percentage Error)分别为0.202,1.198和1.97%,预测性能和其他先进模型相比有明显的提升.证明该模型对复杂气象条件下的污染物预测更准确,并具有较好的泛化性能.Deep learning uses air pollution monitoring data as a standard data set because of its huge number and characteristics of long time series. Aiming at the problems that the existing air pollution prediction methods fail to combine the physical mechanism of atmospheric transport and effectively consider the spatio-temporal characteristics of pollutant transport,in this paper,a feature and time series pollutant prediction model based on multi-model is proposed. Backward air mass trajectories are calculated using HYSPLIT(Hybrid Single-Particle Lagrangian Integrated Trajectory),and the VGG(Convolutional Neural Networks) is introduced to extract the change characteristics of the air mass trajectory. The characteristics is combined with pollutant and meteorological time series data as the input of LSTM(Long Short-Term Memory),then the predicted value of target pollutant concentration in the study area is obtained to evaluate the air quality of the study area. Based on the pollutants and related meteorological online monitoring data from 61 air quality monitoring stations in Guilin,the model performance is evaluated. Experimental results are compared with several state-of-the-art methods,which show that the proposed HYSPLIT-VGG-LSTM(H-VGG-LSTM) model effectively improves the prediction accuracy of atmospheric pollutants. RMSE(The Root Mean Square Error),MAE(Mean Absolute Error),and SMAPE(Symmetric Mean Absolute Percentage Error) of the prediction results are 0.202,1.198,and 1.97% respectively,and the prediction performance of the model is significantly improved compared with other advanced models. It is proved that the proposed model is more accurate in forecasting pollutants for complex meteorological conditions and has better generalization performance.

关 键 词：大气污染预测 PM_(2.5) 后向轨迹模拟 卷积神经网络 LSTM 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]