基于多机器学习模型的逐小时PM2.5浓度预测对比

doi:10.12302/j.issn.1000-2006.202106023

国家林草科技领军期刊
中国精品科技期刊
中国高校百佳科技期刊
江苏省新闻出版政府奖期刊奖
RCCSE林学权威期刊（A+）
CSCD核心期刊
Scopus数据库收录期刊
中文核心期刊
SCD核心期刊

作者加群：102861116

微信公众号：南京林业大学学报

高级检索

南京林业大学学报（自然科学版） ›› 2022, Vol. 46 ›› Issue (5): 152-160.doi: 10.12302/j.issn.1000-2006.202106023

基于多机器学习模型的逐小时PM_2.5浓度预测对比

陈建坤(), 牟凤云(), 张用川, 田甜, 王俊秀

重庆交通大学智慧城市学院,重庆 400074

收稿日期:2021-06-19 修回日期:2021-09-05 出版日期:2022-09-30 发布日期:2022-10-19
通讯作者: 牟凤云
基金资助:
国家重点研发计划(2019YFB2102503);重庆市自然科学基金项目(cstc2019jcyj-msxmX0626)

Comparative analysis of hourly PM_2.5 prediction based on multiple machine learning models

CHEN Jiankun(), MU Fengyun(), ZHANG Yongchuan, TIAN Tian, WANG Junxiu

School of Smart City, Chongqing Jiaotong University, Chongqing 400074,China

Received:2021-06-19 Revised:2021-09-05 Online:2022-09-30 Published:2022-10-19
Contact: MU Fengyun

摘要/Abstract

摘要：

【目的】比较分析XGBoost模型、LightGBM模型、随机森林模型(RF)、K最近邻模型(KNN)、长短期记忆神经网络(LSTM)、决策树模型(DT)共6个PM_2.5浓度预测模型,以准确、及时预测环境PM_2.5浓度。【方法】基于重庆市合川区2020年全年空气质量监测数据和气象数据,通过最大相关最小冗余算法(MRMR)进行数据降维选择最优特征子集,作为模型的输入,逐一进行PM_2.5浓度预测;考虑到不同季节PM_2.5浓度差异较大,故分季节预测了PM_2.5浓度;为了探究各模型预测性能,计算了各模型运行时间和内存占用,并基于PM_2.5与特征变量的相关性和特征变量的重要性探讨了模型预测性能季节性差异原因。【结果】模型总体预测精度从高到低排序为 XGBoost、RF、LightGBM、LSTM、KNN、DT模型;预测性能方面,6个模型均表现为秋冬季节预测精度高于春夏季节;LightGBM模型可在保证模型精度的情况下,大幅减少模型训练时间和内存占用;特征重要性显示PM₁₀浓度、气温和气压的重要性高,O₃浓度、风向和NO₂浓度重要性相对较弱。【结论】采取MRMR方法进行数据降维选取的最优特征子集能较好地预测PM_2.5浓度;相比较而言,XGBoost、RF、LightGBM、LSTM模型在PM_2.5浓度预测上具有较优性能,其中综合性能较好的为LightGBM模型。

关键词: PM_2.5预测, 机器学习, 最大相关最小冗余(MRMR), 气象因子

Abstract:

【Objective】Comparative analysis of the XGBoost model, LightGBM model, random forest model (RF), K nearest neighbor model (KNN), long short-term memory neural network (LSTM), and the decision tree model (DT), a total of six PM_2.5 concentration prediction models was undertaken to ensure accurate and timely prediction of the ambient PM_2.5 concentration.【Method】Based on a full-year of air quality monitoring data and the meteorological data of Hechuan District, Chongqing City in 2020, the maximum minimum redundancy algorithm (MRMR) was used to reduce the data dimensionality to select the optimal feature subset, which is used as the model input. The PM_2.5 concentration prediction was then undertaken one at a time. Considering that the PM_2.5 concentration varies considerably during different seasons, the PM_2.5 concentration was predicted according to season. This was undertaken to explore the prediction performance of each model and the running time and memory usage of each model were calculated. Based on the correlation between PM_2.5 and the characteristic variables and the importance of the characteristic variables, the causes of the seasonal differences in model prediction performance are discussed.【Result】The overall prediction accuracy of the model is ranked as XGBoost, RF, LightGBM, LSTM, KNN, and the DT models. In terms of the prediction performance, the six models all show that the prediction accuracy in autumn and winter is higher than that of spring and summer. The LightGBM model can considerably reduce the training time and memory occupation of the model while ensuring the model accuracy. The importance of these features shows that the importance of PM₁₀, temperature, and the air pressure is high, while the importance of O₃, wind direction, and NO₂ is relatively weak.【Conclusion】The optimal feature subset selected using the MRMR method for data dimensionality reduction can better predict the PM_2.5 concentration. In comparison, the XGBoost, RF, LightGBM, and the LSTM models have higher performance in PM_2.5 prediction, among them, the Light GBM has better comprehensive performance.

Key words: PM_2.5 prediction, machine learning, maximum correlation minimum redundancy (MRMR), meteorological factors

中图分类号:

X513

陈建坤,牟凤云,张用川,等. 基于多机器学习模型的逐小时PM_2.5浓度预测对比[J]. 南京林业大学学报（自然科学版）, 2022, 46(5): 152-160.

CHEN Jiankun, MU Fengyun, ZHANG Yongchuan, TIAN Tian, WANG Junxiu. Comparative analysis of hourly PM_2.5 prediction based on multiple machine learning models[J].Journal of Nanjing Forestry University (Natural Science Edition), 2022, 46(5): 152-160.DOI: 10.12302/j.issn.1000-2006.202106023.

图/表 11

表1

站点逐小时监测数据样本"

数据时间 time	质量浓度/(μg·m^-3) content						气温/℃ air temperature	湿度/% humidity	风速/ (m·s^-1) wind speed	风向/(°) wind direction	气压/kPa air pressure
数据时间 time	PM_2.5	PM₁₀	CO	SO₂	NO₂	O₃	气温/℃ air temperature	湿度/% humidity	风速/ (m·s^-1) wind speed	风向/(°) wind direction	气压/kPa air pressure
2020-01-23 00:00:00	86	116	1.256	20	34	26	12.57	77.83	0.19	4	99.360 98
2020-01-23 01:00:00	88	119	1.350	22	48	10	12.67	79.87	0.15	11	99.364 36
2020-01-23 02:00:00	87	118	1.292	20	39	10	12.24	81.78	0.22	40	99.356 69
2020-01-23 03:00:00	86	116	1.230	19	35	13	12.12	81.43	0.16	36	99.355 83
2020-01-23 04:00:00	79	109	1.194	18	26	23	11.85	83.48	0.18	30	99.311 74
2020-01-23 05:00:00	92	115	1.194	19	28	18	11.79	83.86	0.18	67	99.277 15
2020-01-23 06:00:00	81	108	1.166	19	27	16	11.64	85.29	0.16	48	99.220 48
2020-01-23 07:00:00	82	109	1.200	18	32	11	11.58	85.88	0.23	31	99.237 00
2020-01-23 08:00:00	79	102	1.202	19	35	5	11.59	85.55	0.20	50	99.308 14
2020-01-23 09:00:00	80	98	1.248	17	30	9	11.60	86.42	0.20	83	99.375 48
2020-01-23 10:00:00	88	105	1.253	17	28	11	11.79	86.28	0.20	112	99.445 52
2020-01-23 11:00:00	95	121	1.375	16	33	12	12.48	84.48	0.27	158	99.490 83
2020-01-23 12:00:00	97	123	1.385	15	39	10	12.71	83.67	0.26	159	99.495 35
2020-01-23 13:00:00	99	125	1.330	15	32	19	13.29	81.55	0.43	136	99.416 11
2020-01-23 14:00:00	109	142	1.399	14	40	18	14.24	76.73	0.27	286	99.329 84
2020-01-23 15:00:00	101	135	1.332	14	29	38	14.78	72.54	0.34	296	99.249 16
2020-01-23 16:00:00	83	120	1.245	12	25	60	13.74	75.04	0.43	292	99.227 36
2020-01-23 17:00:00	92	133	1.272	14	27	58	13.57	76.78	0.32	295	99.229 41
2020-01-23 18:00:00	58	90	1.232	11	26	61	13.32	75.50	0.27	298	99.303 92
2020-01-23 19:00:00	89	132	1.299	13	35	50	12.58	80.99	0.39	256	99.384 96
2020-01-23 20:00:00	107	139	1.358	15	41	34	11.89	84.82	0.25	333	99.472 48
2020-01-23 21:00:00	113	142	1.417	16	50	16	11.32	87.04	0.24	11	99.582 50
2020-01-23 22:00:00	114	144	1.330	17	43	14	10.98	89.70	0.18	18	99.644 05
2020-01-23 23:00:00	113	140	1.293	16	36	11	10.80	90.43	0.14	13	99.655 27

表1

表2

表3

图1

图2

图3

表4

表5

图4

表6

图5

参考文献 34

[1]	FUZZI S, BALTENSPERGER U, CARSLAW K, et al. Particulate matter,air quality and climate:lessons learned and future needs[J]. Atmos Chem Phys, 2015, 15(14):8217-8299.DOI:10.5194/acp-15-8217-2015.
[2]	CESARI D, DE BENEDETTO G E, BONASONI P, et al. Seasonal variability of PM_2.5 and PM₁₀ composition and sources in an urban background site in southern Italy[J]. Sci Total Environ, 2018, 612:202-213.DOI:10.1016/j.scitotenv.2017.08.230.
[3]	MANISALIDIS I, STAVROPOULOU E, STAVROPOULOS A, et al. Environmental and health impacts of air pollution:a review[J]. Front Public Health, 2020, 8:14.DOI:10.3389/fpubh.2020.00014.
[4]	KIM K H, KABIR E, KABIR S. A review on the human health impact of airborne particulate matter[J]. Environ Int, 2015, 74:136-143.DOI:10.1016/j.envint.2014.10.005.
[5]	LIU H Y, DUNEA D, IORDACHE S, et al. A review of airborne particulate matter effects on young children's respiratory symptoms and diseases[J]. Atmosphere, 2018, 9(4):150.DOI:10.3390/atmos9040150.
[6]	CHOI S, KIM K H, KIM K, et al. Association between post-diagnosis particulate matter exposure among 5-year cancer survivors and cardiovascular disease risk in three metropolitan areas from south Korea[J]. Int J Environ Res Public Health, 2020, 17(8):2841.DOI:10.3390/ijerph17082841.
[7]	ATKINSON R W, KANG S, ANDERSON H R, et al. Epidemiological time series studies of PM_2.5 and daily mortality and hospital admissions:a systematic review and meta-analysis[J]. Thorax, 2014, 69(7):660-665.DOI:10.1136/thoraxjnl-2013-204492.
[8]	ZHANG Y, BOCQUET M, MALLET V, et al. Real-time air quality forecasting,part I: history,techniques,and current status[J]. Atmos Environ, 2012, 60:632-655.DOI:10.1016/j.atmosenv.2012.06.031.
[9]	李锋, 朱彬, 安俊岭, 等. 2013年12月初长江三角洲及周边地区重霾污染的数值模拟[J]. 中国环境科学, 2015, 35(7):1965-1974.
	LI F, ZHU B, AN J L, et al. Modeling study of a severe haze episode occurred over the Yangtze River Delta and its surrounding regions during early December,2013[J]. China Environ Sci, 2015, 35(7):1965-1974.DOI:10.3969/j.issn.1000-6923.2015.07.008.
[10]	周广强, 谢英, 吴剑斌, 等. 基于WRF-Chem模式的华东区域PM_2.5预报及偏差原因[J]. 中国环境科学, 2016, 36(8):2251-2259.
	ZHOU G Q, XIE Y, WU J B, et al. WRF-Chem based PM_2.5 forecast and bias analysis over the East China region[J]. China Environ Sci, 2016, 36(8):2251-2259.DOI:10.3969/j.issn.1000-6923.2016.08.002.
[11]	DENNIS R L, BYUN D W, NOVAK J H, et al. The next generation of integrated air quality modeling:EPA's models-3[J]. Atmos Environ, 1996, 30(12):1925-1938.DOI:10.1016/1352-2310(95)00174-3.
[12]	CHEN Q Q, TAYLOR D. Transboundary atmospheric pollution in southeast Asia:current methods,limitations and future developments[J]. Crit Rev Environ Sci Technol, 2018, 48(16/17/18):997-1029.DOI:10.1080/10643389.2018.1493337.
[13]	SHIMADERA H, KOJIMA T, KONDO A. Evaluation of air quality model performance for simulating long-range transport and local pollution of PM_2.5 in Japan[J]. Adv Meteorol, 2016, 2016:5694251.DOI:10.1155/2016/5694251.
[14]	郑毅, 朱成璋. 基于深度信念网络的PM_2.5预测[J]. 山东大学学报(工学版), 2014, 44(6):19-25.
	ZHENG Y, ZHU C Z. A prediction method of atmospheric PM_2.5 based on DBNs[J]. J Shandong Univ (Eng Sci), 2014, 44(6):19-25.DOI:10.6040/j.issn.1672-3961.1.2014.180.
[15]	曲悦, 钱旭, 宋洪庆, 等. 基于机器学习的北京市PM_2.5浓度预测模型及模拟分析[J]. 工程科学学报, 2019, 41(3):401-407.
	QU Y, QIAN X, SONG H Q, et al. Machine-learning-based model and simulation analysis of PM_2.5 concentration prediction in Beijing[J]. Chin J Eng, 2019, 41(3):401-407.DOI:10.13374/j.issn2095-9389.2019.03.014.
[16]	李建新, 刘小生, 刘静, 等. 基于MRMR-HK-SVM模型的PM_2.5浓度预测[J]. 中国环境科学, 2019, 39(6):2304-2310.
	LI J X, LIU X S, LIU J, et al. Prediction of PM_2.5 concentration based on MRMR-HK-SVM model[J]. China Environ Sci, 2019, 39(6):2304-2310. DOI:10.3969/j.issn.1000-6923.2019.06.009.
[17]	宋国君, 国潇丹, 杨啸, 等. 沈阳市PM_2.5浓度ARIMA-SVM组合预测研究[J]. 中国环境科学, 2018, 38(11):4031-4039.
	SONG G J, GUO X D, YANG X, et al. ARIMA-SVM combination prediction of PM_2.5 concentration in Shenyang[J]. China Environ Sci, 2018, 38(11):4031-4039.DOI:10.3969/j.issn.1000-6923.2018.11.005.
[18]	康俊锋, 黄烈星, 张春艳, 等. 多机器学习模型下逐小时PM_2.5预测及对比分析[J]. 中国环境科学, 2020, 40(5):1895-1905.
	KANG J F, HUANG L X, ZHANG C Y, et al. Hourly PM_2.5 prediction and its comparative analysis under multi-machine learning model[J]. China Environ Sci, 2020, 40(5):1895-1905.DOI:10.19674/j.cnki.issn1000-6923.2020.0213.
[19]	KAYES I, SHAHRIAR S A, HASAN K, et al. The relationships between meteorological parameters and air pollutants in an urban environment[J]. Glob J Environ Sci Manag, 2019, 5(3):265-278.DOI:10.22034/GJESM.2019.03.01.
[20]	王黎明, 吴香华, 赵天良, 等. 基于距离相关系数和支持向量机回归的PM_2.5浓度滚动统计预报方案[J]. 环境科学学报, 2017, 37(4):1268-1276.
	WANG L M, WU X H, ZHAO T L, et al. A scheme for rolling statistical forecasting of PM_2.5 concentrations based on distance correlation coefficient and support vector regression[J]. Acta Sci Circumstantiae, 2017, 37(4):1268-1276.DOI:10.13671/j.hjkxxb.2016.0345.
[21]	JUHOS I, MAKRA L, TÓTH B. Forecasting of traffic origin NO and NO₂ concentrations by support vector machines and neural networks using principal component analysis[J]. Simul Model Pract Theory, 2008, 16(9):1488-1502.DOI:10.1016/j.simpat.2008.08.006.
[22]	王占山, 李云婷, 陈添, 等. 2013年北京市PM_2.5的时空分布[J]. 地理学报, 2015, 70(1):110-120.
	WANG Z S, LI Y T, CHEN T, et al. Spatial-temporal characteristics of PM_2.5 in Beijing in 2013[J]. Acta Geogr Sin, 2015, 70(1):110-120.DOI:10.11821/dlxb201501009.
[23]	郭立力, 赵春江. 十折交叉检验的支持向量机参数优化算法[J]. 计算机工程与应用, 2009, 45(8):55-57.
	GUO L L, ZHAO C J. Optimizing parameters of support vector machine's model based on genetic algorithm[J]. Comput Eng Appl, 2009, 45(8):55-57.DOI:10.3778/j.issn.1002-8331.2009.08.017.
[24]	CHEN T, TONG H, BENESTY M. Xgboost: eXtreme gradient boosting[M]. London: Sage Publications, 2016:931-961.
[25]	方匡南, 吴见彬, 朱建平, 等. 随机森林方法研究综述[J]. 统计与信息论坛, 2011, 26(3):32-38.
	FANG K N, WU J B, ZHU J P, et al. A review of technologies on random forests[J]. Stat Inf Forum, 2011, 26(3):32-38.DOI:10.3969/j.issn.1007-3116.2011.03.006.
[26]	KE G, MENG Q, FINLEY T, et al. LightGBM: a highly efficient gradient boosting decision tree[C]// Proceeding of the 13th interenational Conference Neural Information Processing Systems, New York: ACM, 2017:3149-3157.
[27]	桑应宾. 基于K近邻的分类算法研究[D]. 重庆: 重庆大学, 2009.
	SANG Y B. Research of classification algorithm based on K nearest neighbor[D]. Chongqing: Chongqing University, 2009.
[28]	CARRERA-GARCÍA L, MUCHART J, LAZARO J J, et al. Pediatric SMA patients with complex spinal anatomy:implementation and evaluation of a decision-tree algorithm for administration of nusinersen[J]. Eur J Paediatr Neurol, 2021, 31:92-101.DOI:10.1016/j.ejpn.2021.02.009.
[29]	LI X, PENG L, YAO X J, et al. Long short-term memory neural network for air pollutant concentration predictions:method development and evaluation[J]. Environ Pollut, 2017, 231(Pt 1):997-1004.DOI:10.1016/j.envpol.2017.08.114.
[30]	PENG H C, LONG F H, DING C. Feature selection based on mutual information:criteria of max-dependency,max-relevance,and min-redundancy[J]. IEEE Trans Pattern Anal Mach Intell, 2005, 27(8):1226-1238.DOI:10.1109/TPAMI.2005.159.
[31]	BAE JE, CHOI H, SHIN D W, et al. Fine particulate matter (PM_2.5) inhibits ciliogenesis by increasing SPRR3 expression via c-Jun activation in RPE cells and skin keratinocytes[J]. Sci Rep, 2019, 9(1):3994.DOI:10.1038/s41598-019-40670-y.
[32]	环境保护部. HJ 633-2012:环境空气质量指数(AQI)技术规定(试行)[EB/OL]. [2021-05-10]. http://www.mee.gov.cn/ywgz/fgbz/bz/bzwb/jcffbz/201203/W020120410332725219541.pdf.
[33]	MA X Y, JIA H L, SHA T, et al. Spatial and seasonal characteristics of particulate matter and gaseous pollution in China: implications for control policy[J]. Environ Pollut, 2019, 248:421-428.DOI:10.1016/j.envpol.2019.02.038.
[34]	曾昭亮, 郭建平, 马大喜. 基于江西地区多卫星数据的气溶胶立体分布研究[J]. 大气与环境光学学报, 2016, 11(5):391-400.
	ZENG Z L, GUO J P, MA D X. Research of aerosol three-dimensional distribution based on multi-satellite data over Jiangxi[J]. J Atmos Environ Opt, 2016, 11(5):391-400.DOI:10.3969/j.issn.1673-6141.2016.05.007.

模型 model	模型特点 model characteristic	参数确定 parameter determination
XGBoost	基于决策树的集成机器学习算法,以梯度提升(gradient boost)为框架,由多个弱分类器集成而构建强分类器^[24]	n_estimators=300,ooster="dart", max_depth=9,learning_rate=0.1,reg_lambda=0.2
随机森林 RF	基于分类树算法,通过对大量分类树的汇总提高预测精度,对于异常值和噪声具有很好的容忍度,且预测时不易出现过拟合,是一种非线性建模的工具^[25]	n_estimators=600,oob_score=True
LightGBM	一种梯度提升决策树框架,具有训练速度快、效率高、内存占用低、准确性高、支持并行和GPU学习、能够处理大规模数据等优点^[26]	max_depth=9,num_leaves=30, ubsample=0.5,learning_rate=0.1, min_data_in_leaf=21
近邻算法 KNN	近邻算法是将数据集合中每一个记录进行分类的方法。数据预测时KNN对于给定的输入X,在历史输入中搜索找出距离最近的K个特征值,然后对K个特征值进行加权估计即可得到预测值^[27]	K=11,weights = ‘distance’
模型 model	模型特点 model characteristic	参数确定 parameter determination
决策树 DT	用于分类和回归任务的机器学习算法,可选择一个或多个变量作为输入变量,仅有单一输出,具有训练时间快、内存占用小的特点^[28]	max_depth: 19, min_samples_leaf: 12, min_samples_split: 6
长短期记忆神经网络LSTM	特殊的RNNs,网络结构中重复的单元不用,重复的单元被称为memory block(记忆块)。主要包含了3个门(forget gate、input gate、output gate)与1个记忆单元(cell)。网络中的cell state(单元状态)可以控制信息传递给下一时刻。在解决非线性时间序列数据时优势明显^[29]	output_dim=60,activation=‘relu’, epochs=30, batch_size = 72,loss=‘mae’, optimizer=‘adam’

等级 grade	实测值 measured value	XGBoost	RF	LightGBM	DT	KNN	LSTM
1	6 285	6 224	6 248	6 220	6 230	6 343	6 099
2	4 260	4 275	4 281	4 301	4 245	4 221	4 333
3	1 451	1 548	1 502	1 525	1 564	1 496	1 585
4	401	378	390	374	375	381	391
5	146	121	123	122	130	114	136
6	13	10	12	14	12	11	12

模型 model	春季spring				夏季summer				秋季autumn				冬季winter
模型 model	R²	σ_RMSE	σ_MAE	σ_MAPE/ %	R²	σ_RMSE	σ_MAE	σ_MAPE/ %	R²	σ_RMSE	σ_MAE	σ_MAPE/ %	R²	σ_RMSE	E_MAE	σ_MAPE/ %
XGBoost	0.905	9.127	5.479	14.852	0.900	6.364	4.107	21.752	0.922	8.367	4.680	16.131	0.948	8.085	5.720	9.697
RF	0.901	9.298	5.573	15.247	0.906	6.174	4.066	21.736	0.915	8.718	4.823	16.636	0.946	8.267	5.952	10.041
LightGBM	0.903	9.214	5.754	16.440	0.902	6.295	4.156	22.912	0.915	8.717	4.952	17.790	0.942	8.567	6.096	10.339
DT	0.861	11.019	6.453	17.430	0.876	7.098	4.681	24.701	0.895	9.687	5.451	18.936	0.927	9.585	6.848	11.441
KNN	0.890	9.805	6.157	17.669	0.889	6.693	4.472	25.308	0.905	9.210	5.322	19.870	0.939	8.753	6.311	11.238
LSTM	0.889	9.845	6.028	17.639	0.885	6.822	4.455	24.485	0.932	7.816	5.033	18.477	0.937	8.877	6.342	10.949

模型 model	占用内存大小/GB memory footprint	运行时间/s running time
XGBoost	0.267 2	22.3
LightGBM	0.200 4	10.9
RF	2.071 4	178.0
DT	0.193 0	10.2
KNN	0.209 0	14.3
LSTM	0.237 0	46.5