基于深度时间序列特征融合的西安市2015—2020年供暖季雾霾重污染过程预警

doi:10.16085/j.issn.1000-6613.2021-2658

摘要/Abstract

摘要：

为准确预测雾霾重污染演变过程，本文提出深度时间序列特征融合模型（long short-term memory and multivariate linear regression，LSTM-MLR）并对西安市PM_2.5浓度进行了临近预测。该模型利用不同超参数长短期记忆网络（long short-term memory，LSTM）提取PM_2.5前体和气象因素时间序列中的深度特征；采用多元线性回归（multivariate linear regression，MLR）形式融合LSTM单元输出的深度时间序列特征，最终输出PM_2.5浓度预测值。为评估模型性能，采用西安市2015年1月至2020年3月采暖季数据进行建模并计算未来3h、6h、12h、24h的PM_2.5浓度预测精度。结果表明：LSTM-MLR模型对雾霾严重污染样本的准确预测率分别为94.12%、85.29%、77.57%和51.10%，显著高于随机森林（random forest，RF）、支持向量回归（support vector regression，SVR）、MLR、单变量LSTM（LSTM_PM_2.5）、多变量LSTM（M_LSTM）和LSTM-RF（long short-term memory and random forest）；融合系数显示当前PM_2.5浓度对未来PM_2.5浓度的影响随预测步长的增加从80.89%（t+3）急剧降低至16.34%（t+24），前体浓度影响力从5.23%（t+3）上升至29.43%（t+24），说明提前控制前体物排放强度对雾霾重污染事件消峰降速效果具有显著影响。

关键词: 雾霾重污染, LSTM-MLR模型, 预测, 多尺度, 神经网络

Abstract:

To achieve accurate prediction of PM_2.5 concentrations during heavy haze pollution events, the model of deep feature fusion (long short-term memory and multivariate linear regression, LSTM-MLR) was proposed in this paper to predict PM_2.5 concentrations in 3h, 6h, 12h and 24h of Xi’an and to explain the relationship between the main meteorological factors precursors and haze concentration. In the proposed model, a series of long short-term memory (LSTM) with different hyper-parameter were used to extract the deep features of PM_2.5 precursors and meteorological time series, and a full connect layer was used to adjust the dimensions of LSTM outputs. Then, a feature fusion model with the structure of multivariable linear regression (MLR) was used to obtain the predicted PM_2.5 concentrations. Data from January 2015 to March 2020 in Xi’an were used to determine parameters of the proposed model, and model performances for PM_2.5 concentration prediction in 3h, 6h, 12h and 24h were also evaluated. Results showed that LSTM-MLR, which was based on deep feature fusion of time series, can predict heavy haze pollution samples correctly with the accuracy of 94.12%, 85.29%, 77.57% and 51.10% in 3h, 6h, 12h and 24h, respectively. The true positive rate (TPR) of the proposed model outperforms random forest (RF), support vector regression (SVR), MLR, LSTM_PM_2.5 (PM_2.5 used only), multivariable LSTM (M_LSTM) and long short-term memory-random forest (LSTM-RF) which had the same input as LSTM-MLR. The fusion coefficients showed that the influence of current PM_2.5 concentrations on that in the future decreased from 80.89% (t + 3) to 16.34% (t + 24), and the influence of precursor concentration increased from 5.23% (t + 3) to 29.43% (t + 24), which illustrated the importance of time for taking emergency measures to reduce the peak value and decrease the pollution duration.

Key words: heavy haze pollution event, LSTM-MLR model, prediction, multiscale, neural networks

中图分类号:

X513

王英, 冉进业, 张今, 杨鑫, 张浩. 基于深度时间序列特征融合的西安市2015—2020年供暖季雾霾重污染过程预警[J]. 化工进展, 2022, 41(10): 5685-5694.

WANG Ying, RAN Jinye, ZHANG Jin, YANG Xin, ZHANG Hao. Prediction of heavy haze pollution episodes based on deep feature fusion of pollutant and meteorological time series in Xi’an during 2015—2020 heating season[J]. Chemical Industry and Engineering Progress, 2022, 41(10): 5685-5694.

图/表 10

参考文献 26

1	LE Tianhao, WANG Yuan, LIU Lang, et al. Unexpected air pollution with marked emission reductions during the COVID-19 outbreak in China[J]. Science, 2020, 369(6504): 702-706.
2	ZHAO Shuyun, ZHANG Hua, XIE Bing. The effects of El Niño-Southern Oscillation on the winter haze pollution of China[J]. Atmospheric Chemistry and Physics, 2018, 18(3): 1863-1877.
3	ZHANG Qiang, ZHENG Yixuan, TONG Dan, et al. Drivers of improved PM_2.5 air quality in China from 2013 to 2017[J]. PNAS, 2019, 116(49): 24463-24469.
4	XU Peng, CHEN Yongfen, YE Xiaojian. Haze, air pollution, and health in China[J]. The Lancet, 2013, 382(9910): 2067.
5	朱彤, 尚静, 赵德峰. 大气复合污染及灰霾形成中非均相化学过程的作用[J]. 中国科学: 化学, 2010, 40(12): 1731-1740.
	ZHU Tong, SHANG Jing, ZHAO Defeng. The roles of heterogeneous chemical processes in the formation of an air pollution complex and gray haze[J]. Scientia Sinica Chimica, 2010, 40(12): 1731-1740.
6	PETÄJÄ T, JÄRVI L, KERMINEN V M, et al. Enhanced air pollution via aerosol-boundary layer feedback in China[J]. Scientific Reports, 2016, 6: 18998.
7	熊远南. 基于改进灰色-多元回归组合预测模型的燃煤电厂智慧水务研究[J]. 化工进展, 2020, 39(S2): 393-400.
	XIONG Yuannan. Smart water affair of coal-fired power plant based on improved combination prediction model of grey system and regression analysis[J]. Chemical Industry and Engineering Progress, 2020, 39(S2): 393-400.
8	GARIAZZO C, CARLINO G, SILIBELLO C, et al. A multi-city air pollution population exposure study: combined use of chemical-transport and random-Forest models with dynamic population data[J]. The Science of the Total Environment, 2020, 724: 138102.
9	靳文博, 敬加强, 田震, 等. 基于最小二乘支持向量机的蜡沉积速率预测[J]. 化工进展, 2014, 33(10): 2565-2569.
	JIN Wenbo, JING Jiaqiang, TIAN Zhen, et al. Prediction of wax deposition rate based on least squares support vector machine[J]. Chemical Industry and Engineering Progress, 2014, 33(10): 2565-2569.
10	张浩, 于君毅, 刘晓慧, 等. 基于广义隐马尔可夫模型的PM_2.5浓度预测[J]. 化工学报, 2018, 69(3): 1215-1220.
	ZHANG Hao, YU Junyi, LIU Xiaohui, et al. Prediction of fine particulate matter concentrations based on generalized hidden Markov model[J]. CIESC Journal, 2018, 69(3): 1215-1220.
11	SUN Wei, ZHANG Hao, PALAZOGLU A, et al. Prediction of 24-hour-average PM_2.5 concentrations using a hidden Markov model with different emission distributions in Northern California[J]. Science of the Total Environment, 2013, 443: 93-103.
12	姚羽曼, 罗文嘉, 戴一阳. 数据驱动方法在化工过程故障诊断中的研究进展[J]. 化工进展, 2021, 40(4): 1755-1764.
	YAO Yuman, LUO Wenjia, DAI Yiyang. Research progress of data-driven methods in fault diagnosis of chemical process[J]. Chemical Industry and Engineering Progress, 2021, 40(4): 1755-1764.
13	宋泓阳, 孙晓岩, 项曙光. 人工神经网络在化工过程中的应用进展[J]. 化工进展, 2016, 35(12): 3755-3762.
	SONG Hongyang, SUN Xiaoyan, XIANG Shuguang. Progress on the application of artificial neural network in chemical industry[J]. Chemical Industry and Engineering Progress, 2016, 35(12): 3755-3762.
14	GIRSHICK R. Fast R-CNN[C]//2015 IEEE International Conference on Computer Vision. IEEE, 2015: 1440-1448.
15	HE Kaiming, GKIOXARI G, DOLLÁR P, et al. Mask R-CNN[C]// 2017 IEEE International Conference on Computer Vision. IEEE, 2017: 2980-2988.
16	SILVER D, SCHRITTWIESER J, SIMONYAN K, et al. Mastering the game of Go without human knowledge[J]. Nature, 2017, 550(7676): 354-359.
17	SENIOR A W, EVANS R, JUMPER J, et al. Improved protein structure prediction using potentials from deep learning[J]. Nature, 2020, 577(7792): 706-710.
18	JUMPER J, EVANS R, PRITZEL A, et al. Highly accurate protein structure prediction with AlphaFold[J]. Nature, 2021, 596(7873): 583-589.
19	ZHAO Jiachen, DENG Fang, CAI Yeyuan, et al. Long short-term memory-fully connected (LSTM-FC) neural network for PM_2.5 concentration prediction[J]. Chemosphere, 2019, 220: 486-492.
20	马双忱, 林宸雨, 周权, 等. 基于深度神经网络的脱硫系统预测模型及应用[J]. 化工进展, 2021, 40(3): 1689-1698.
	MA Shuangchen, LIN Chenyu, ZHOU Quan, et al. Prediction model of FGD system based on deep neural network and its application[J]. Chemical Industry and Engineering Progress, 2021, 40(3): 1689-1698.
21	方黄峰, 刘瑶瑶, 张文彪. 基于LSTM神经网络的流化床干燥器内生物质颗粒湿度预测[J]. 化工学报, 2020, 71(S1): 307-314.
	FANG Huangfeng, LIU Yaoyao, ZHANG Wenbiao. Biomass moisture content prediction in fluidized bed dryer based on LSTM neural network[J]. CIESC Journal, 2020, 71(S1): 307-314.
22	GHORBANI A, OUYANG D, ABID A, et al. Deep learning interpretation of echocardiograms[J]. NPJ Digital Medicine, 2020, 3: 10.
23	DU Mengnan, LIU Ninghao, HU Xia. Techniques for interpretable machine learning[J]. Communications of the ACM, 2019, 63(1): 68-77.
24	GREFF K, SRIVASTAVA R K, KOUTNIK J, et al. LSTM: a search space odyssey[J]. IEEE Transactions on Neural Networks and Learning Systems, 2017, 28(10): 2222-2232.
25	SUN Wei, ZHANG Hao, PALAZOGLU A. Prediction of 8 h-average ozone concentration using a supervised hidden Markov model combined with generalized linear models[J]. Atmospheric Environment, 2013, 81: 199-208.
26	徐丹妮, 王瑾婷, 袁自冰, 等. 汾渭平原复杂地形影响下冬季PM_2.5污染分布特征、来源及成因分析[J]. 环境科学学报, 2021, 41(4): 1184-1198.
	XU Danni, WANG Jinting, YUAN Zibing, et al. Temporal-spatial variations, source apportionment, and formation mechanisms of PM_2.5 pollution over Fenwei Plain, China[J]. Acta Scientiae Circumstantiae, 2021, 41(4): 1184-1198.

变量	+3h		+6h		+12h		+24h
变量	P-value	r	P-value	r	P-value	r	P-value	r
PM_2.5	0	0.98	0	0.93	0	0.84	4.98E-95	0.67
PM₁₀	2.76×10^-7	0.87	3.6×10^-8	0.82	—	—	2.26×10^-44	0.63
CO	—	—	—	—	—	—	—	—
NO₂	3.6×10^-17	0.80	3.93×10^-57	0.78	3.75×10^-14	0.67	3.54×10^-37	0.58
SO₂	—	—	—	—	4.24×10^-8	0.45	9.10×10^-18	0.38
O₃	—	—	—	—	—	—	—	—
p_O	2.71×10^-10	0.32	8.93×10^-13	0.33	—	—	—	—
p	—	—	—	—	8.93×10^-12	0.32	3.9×10^-31	0.29
RH	3.69×10^-8	0.56	—	—	—	—	1.00×10^-5	0.44
WD	7.61×10^-8	0.38	2.43×10^-49	0.35	1.02×10^-67	0.33	3.45×10^-12	0.30
WS	0.0271	0.48	—	—	—	—	7.79×10^-74	0.31
T_d	—	—	—	—	—	—	—	—
Temp	0.0139	0.28	1.96×10^-26	0.34	4.62×10^-90	0.32	1.50×10^-5	0.43
VV	2.04×10^-10	0.69	1.73×10^-31	0.33	1.38×10^-39	0.33	3.86×10^-46	0.24

变量	+3h		+6h		+12h		+24h
变量	P-value	r	P-value	r	P-value	r	P-value	r
PM_2.5	0	0.98	0	0.93	0	0.84	4.98E-95	0.67
PM₁₀	2.76×10^-7	0.87	3.6×10^-8	0.82	—	—	2.26×10^-44	0.63
CO	—	—	—	—	—	—	—	—
NO₂	3.6×10^-17	0.80	3.93×10^-57	0.78	3.75×10^-14	0.67	3.54×10^-37	0.58
SO₂	—	—	—	—	4.24×10^-8	0.45	9.10×10^-18	0.38
O₃	—	—	—	—	—	—	—	—
p_O	2.71×10^-10	0.32	8.93×10^-13	0.33	—	—	—	—
p	—	—	—	—	8.93×10^-12	0.32	3.9×10^-31	0.29
RH	3.69×10^-8	0.56	—	—	—	—	1.00×10^-5	0.44
WD	7.61×10^-8	0.38	2.43×10^-49	0.35	1.02×10^-67	0.33	3.45×10^-12	0.30
WS	0.0271	0.48	—	—	—	—	7.79×10^-74	0.31
T_d	—	—	—	—	—	—	—	—
Temp	0.0139	0.28	1.96×10^-26	0.34	4.62×10^-90	0.32	1.50×10^-5	0.43
VV	2.04×10^-10	0.69	1.73×10^-31	0.33	1.38×10^-39	0.33	3.86×10^-46	0.24

模型	训练时间/s	在线预测速度/ms‧样本^-1
RF	471	187.50
SVR	115	629.35
MLR	1	180.52
LSTM_PM_2.5	593	425.19
M_LSTM	800	660.56
LSTM-RF	1064	639.69
LSTM-MLR	594	605.71

模型	训练时间/s	在线预测速度/ms‧样本^-1
RF	471	187.50
SVR	115	629.35
MLR	1	180.52
LSTM_PM_2.5	593	425.19
M_LSTM	800	660.56
LSTM-RF	1064	639.69
LSTM-MLR	594	605.71

模型	+3h				+6h
模型	MAPE/%	MAE/μg‧m^-3	RMSE/μg‧m^-3	R²	MAPE/%	MAE/μg‧m^-3	RMSE/μg‧m^-3	R²
RF	19.09	14.37	19.58	0.91	32.37	20.65	28.60	0.77
SVR	38.32	19.02	24.04	0.89	37.44	22.10	28.91	0.79
MLR	21.23	13.18	17.66	0.91	32.48	19.21	26.30	0.80
LSTM_PM_2.5	11.79	8.36	11.84	0.96	29.46	16.27	22.67	0.85
M_LSTM	12.62	8.44	11.73	0.96	26.21	16.09	22.92	0.84
LSTM-RF	13.42	8.83	12.28	0.96	26.06	16.25	23.28	0.84
LSTM-MLR	11.94	7.89	11.46	0.96	26.31	15.18	21.37	0.86
模型	+12h				+24h
模型	MAPE/%	MAE/μg‧m^-3	RMSE/μg‧m^-3	R²	MAPE/%	MAE/μg‧m^-3	RMSE/μg‧m^-3	R²
RF	53.10	29.16	38.93	0.56	72.17	36.13	48.54	0.32
SVR	48.81	28.03	36.82	0.62	67.68	33.21	44.72	0.42
MLR	48.57	27.45	36.48	0.62	67.15	32.92	44.47	0.42
LSTM_PM_2.5	50.12	25.39	34.45	0.65	84.52	36.36	48.55	0.35
M_LSTM	41.30	24.01	33.85	0.66	65.55	32.96	45.56	0.40
LSTM-RF	44.69	24.07	32..80	0.69	69.53	34.15	45.15	0.39
LSTM-MLR	46.03	23.07	32.59	0.70	71.80	33.56	45.20	0.42