基于树模型的北京市PM_(2.5)预测效果对比分析  

A COMPARATIVE STUDY ON EDICTIVE EFFECT OF PM_(2.5) IN BEIJING BASED ON TREE MODELS

作　　者：李志生[1] 梁锡冠 金宇凯 张华刚 欧耀春 LI Zhi-sheng;LIANG Xi-guan;JIN Yu-kai;ZHANG Hua-gang;OU Yao-chun(School of Civil and Transportation Engineering,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学土木与交通工程学院,广州510006

出　　处：《环境工程》2021年第6期106-113,共8页Environmental Engineering

基　　金：广东省自然科学基金(S2011040003755);广东工业大学产学研合作项目“工业厂房室内甲醛与PM_(2.5)治理项目”(18HK0031)。

摘　　要：在城市空气质量预测中,ρ(PM_(2.5))会受到气象条件和时间周期的影响。选取北京市全市为实验区域,对多种污染物浓度特征、时间特征及天气特征等进行分析,采用2019年33个空气质量监测站逐小时数据开展PM_(2.5)预测实验,建立了基于特征的LightGBM (light gradient boosting machine) PM_(2.5)质量浓度预测模型,分别与随机森林模型(RF)、梯度提升树模型(GBDT)、 XGBoost模型3个PM_(2.5)浓度预测模型进行对比。结果表明:在PM_(2.5)浓度预测精度方面,LightGBM模型最高,XGBoost模型次之,RF模型最差。LightGBM模型的PM_(2.5)污染浓度预测准确率高于其他模型,R2为0.9614,且具有训练快、内存少等优点。LightGBM模型的5个评估指标均优于其他模型,说明其在PM_(2.5)逐时预测上具有很好的稳定性和应用前景。In urban air quality forecast, the mass concentrations of PM_(2.5) were influenced by the meteorological conditions and time period. This article selected Beijing as the experimental area, analysing a variety of pollutants concentration characteristics, time characteristics and weather characteristics. The data by hour of 33 air quality monitoring stations in 2019 were used to carry out the PM_(2.5) forecast experiments, based on characteristics of LightGBM(light gradient boosting machine) PM_(2.5) mass concentration prediction model. The results showed that compared with random forests model(RF), gradient boosting decision tree model(GBDT), XGBoost model, LightGBM model had the highest prediction accuracy of PM_(2.5) concentration, XGBoost model came next, random forest model was the lowest. The accuracy of LightGBM model PM_(2.5) prediction was higher than other models, R2 was 0.9614, and training LightGBM model was fast and RAM needed less. LightGBM model on the five indicators were better than the rest of the model, and LightGBM model on PM_(2.5) hourly prediction had better stability and application prospects.

关 键 词：周期特征 机器学习 PM_(2.5)影响因素 LightGBM PM_(2.5)预测 

分 类 号：X831[环境科学与工程—环境工程]