main-content

## Weitere Artikel dieser Ausgabe durch Wischen aufrufen

18.06.2018 | S.I. : Emerging Intelligent Algorithms for Edge-of-Things Computing | Ausgabe 5/2019 Open Access

# Study on the spatial–temporal change characteristics and influence factors of fog and haze pollution based on GAM

Zeitschrift:
Neural Computing and Applications > Ausgabe 5/2019
Autoren:
Zhuang Wu, Shuo Zhang

## 1 Introduction

Haze, as a collective name for fog and haze, is a kind of weather phenomenon in meteorological science. Fog is the outcome of condensation of water vapor (or deposition) in the air of atmospheric surface layer, and the core materials of haze include the aerosol particulate matter suspended in the air, mainly coming from such artificial sources as industrial pollution, fossil fuel combustion and biomass burning as well as the natural sources like soil dust. PM2.5 is not only an important component of haze, but also a key object for air quality monitoring. It degrades the atmospheric visibility and increases the morbidity and mortality of respiratory disease and cerebrovascular disease [1, 2]. In recent years, unusual hazy weather has been frequently sweeping over China, dramatically deteriorating air quality and severely affecting normal production activities. Since the year of 2012, PM2.5 has been added to “Air Quality Standards” as a conventional index and its real-time concentration has also been appended to the air quality monitoring system of Ministry of Environmental Protection and the People’s Republic of China. Therefore, to figure out the related influence factors and diffusion process of PM2.5 is vital to find an effective governance approach.
Existing international study on fog and haze mainly includes: (1) the composition and sources of haze pollutants, (2) the correlation between urban air pollution and meteorological elements and (3) the virtual and forecasting of numeric value of regional airborne pollutant concentration. With regard to the root of haze, Zhang et al. [3] have believed that the use of non-clean energy is the most fundamental cause of the pollution of fog and haze in China and that the exploding urbanization process in China in recent years is the immediate cause [4]. Related meteorological conditions, including temperature, relative humidity and height of planetary boundary layer, play a significant role in the formation, distribution, maintenance and variation of these pollutants. Most previous study on influence factors and prediction of PM2.5 concentration has adopted the method of linear regression, but in fact, there is complicated relationship, usually nonlinear, between these factors. So, scholars have begun to use generalized additive model to depict the complex relations between these potential variables and PM2.5 concentration in recent years. Song and others [5] have applied generalized additive model and multi-source monitoring data to describe and changes of PM2.5 concentration in Xi’an City. For the independent variables, SO2 and CO will be considered as linear functions, while NO2, O3, AOT and temperature as single-variable smooth nonlinear functions and wind scale as two-variable smooth nonlinear function. The final explanation rate of the model is 69.50%. Li et al. [6] have brought in principal component analysis (PCA) approach into generalized additive model (GAM), proposed and used PCA-GAM to forecast PM2.5 concentrations in Beijing Municipality, Tianjin Municipality and Hebei Province, and the results have shown that compared with conventional land-use regression model, that model has a higher accuracy rate (R2 = 0.94). In conclusion, PM2.5 concentration increases for the following two causes: the meteorological elements featured by southeast wind, temperature inversion and high relative humidity which go against diffusion and the pollution factors featured by the increase in the concentration of suspended particulate matter [712]. Among them, pollution is the internal cause, which is closely related to human activities and which can be controlled while meteorological element is the external cause and it cannot be controlled [13].
To sum up, the foregoing research work mainly integrates specific pollution process and analyzes the numeric changes of PM2.5 and the impact a single meteorological factor has on PM2.5; in other words, it only conducts research on fog and haze within a short time, and it is not universally applicable. Even for the same city, PM2.5, together with air pollutants, meteorological elements and other factors, has formed a complex nonlinear dynamic system, which has multilayered scale structure and local variations in the time domain [14]. Therefore, based on the analysis of change rule of particulate matter (PM2.5), the promotion or inhibition that precursor pollutants (i.e., SO2, NO2, CO) and meteorological elements have on PM2.5 has been taken into consideration, and research has been conducted to PM2.5 concentration in Beijing from the dimensions of time and space by using the air quality data released by Beijing Municipal Environmental Monitoring Center and the meteorological data published by China Meteorological Administration to analyze its features in different stages. Generalized additive model (GAM) is a flexible statistical model, which can be used to detect the influence of nonlinear regression. Nonparametric regression does not require models to satisfy linear assumptions and can detect the complex relationship between data flexibly. However, when the number of independent variables in the model is more, the estimated variance of the model will increase, and each additive item in the GAM model is estimated by a single smooth function. In each additive item, how the dependent variable varies with the independent variable can be explained well. GAM is constructed with the focus on the analysis of the impact the interaction of different influence factors has on the changes of PM2.5 concentration to find the key influence factors and analyze the entire change process of PM2.5 concentration.

## 2 Data and research methods

### 2.1 Data

According to the information provided by Beijing Municipal Environmental Monitoring Center, 35 air pollution monitoring stations have been selected with complete coverage of all districts of Beijing, and the information and locations of every monitoring station are shown in Fig. 1. Collect daily and hourly data of SO2 (ug/m3), NO2 (ug/m3), CO (ug/m3), O3 (ug/m3) and PM2.5 (ug/m3), respectively, and the meteorological data, including wind speed (km/h), relative humidity (%), temperature (°C), atmospheric pressure (hpa), precipitation (mm) and sunshine hour (h), from December 1, 2016, to November 30, 2017. Among them, air quality data all come from Beijing Municipal Environmental Monitoring Center (www.​bjmemc.​com.​cn) and meteorological data from China Meteorological Administration (www.​cma.​gov.​cn).

### 2.2 Research methods

#### 2.2.1 Analysis of spatial and temporal characteristics of PM2.5

Based on the air quality monitoring data of 16 districts of Beijing during December 2016 and November 2017, such methods as time-series analysis, time-series plot and spatial correlation analysis have been adopted to study the rule of temporal and spatial variation of PM2.5 concentration, and Kruskal–Wallis test, Mann–Whitney U test and Bonferroni correction have been conducted on PM2.5 concentration, respectively, to clarify their differences. By reference to the air quality standards released by environmental protection departments, the air quality is excellent when PM2.5 is within 0–35 ug/m3, good when within 35–75 ug/m3, slight (intermediate) pollution when within 75–150 ug/m3 and heavy pollution when bigger than 150 ug/m3. In order to simplify the research process, 16 districts of Beijing have been divided into three regions from north to south (see Table 1) in four seasons: spring (March–May), summer (June–August), autumn (September–November) and winter (December–February). Besides, time-series plot is used to depict the changes in daily PM2.5 concentration.
Table 1
Distribution of PM2.5 concentrations in the 16 districts of Beijing, 2016–2017 (ug/m3)
Regional category
District
Mean
SD
Rate I (%)
Rate II (%)
North
Yanqing
52.92
40.74
4.81–308.80
191 (55.20)
Huairou
56.23
56.08
5.42–485.00
187 (54.05)
Miyun
57.35
56.74
4.86–459.50
195 (56.36)
Shunyi
65.30
63.51
4.43–398.00
216 (62.42)
Changping
56.97
56.84
4.71–413.75
188 (54.17)
Pinggu
70.04
65.80
6.07–419.75
239 (69.08)
Center
Dongcheng
71.73
69.91
7.38–430.92
237 (68.59)
Haidian
65.32
65.09
5.38–463.50
219 (63.29)
Xicheng
72.49
68.65
6.88–462.00
239 (69.08)
Chaoyang
72.85
66.29
5.92–406.67
255 (73.70)
ShiJingShan
71.15
71.25
4.83–495.00
233 (67.34)
Fengtai
77.77
71.10
6.83–495.13
277 (80.06)
South
Mentougou
65.61
66.37
6.04–551.00
216 (62.43)
Touzhou
85.21
76.40
6.98–458.46
278 (80.35)
Daxing
79.30
73.58
7.76–443.03
260 (75.14)
Fangshan
91.13
77.68
6.78–502.31
290 (83.81)
SD standard deviation
Rate I (%): non-attainment days and rates (%) for grade I (0–35 ug/m3)
Rate II (%): non-attainment days and rates (%) for grade II (35–75 ug/m3)

The function in GAM can be identified by the reverse fitting method; it can be applied to the analysis of a variety of distributed data. The model can include both the parameter fitting part and the nonparametric fitting part. All the interpretative components in the GAM are all kinds of smoothing function forms to explain the variables. It is suitable for the analysis of many complex linear relations. In recent years, nonparametric models have attracted increasing attention from scholars. Hastie and Tibshirani have applied additive model into generalized linear model (GLM) and come up with the concept of generalized additive model (GAM) in 1990, which, in its essence, is to connect the expectation of response variable with the additivity in the additive model via the connection function [15] in the following formula.
$$g(E(Y)) = s_{0} + s_{1} (X_{1} ) + s_{2} (X_{2} ) + \cdots s_{p} (X_{p} )$$
(1)
In this formula, E(Y) is the expectation of Y. s0 is the intercept and si (.) (i = 1, 2,…p) is a nonparametric smooth function and Esi(Xi) = 0. It can be a smooth spline function, a local regression smooth function or a kernel function. g(.) is the connection function and g(.) can be represented in the following forms for predictive variables of different distribution types.
$$g(E(Y)) = \left\{ {\begin{array}{*{20}l} {E(Y)} \hfill & {Y\sim\,N(\mu , \, \sigma^{2} )} \hfill \\ {\log \left[ {\frac{E(Y)}{1 - E(Y)}} \right]} \hfill & {Y\sim\,B \, (n, \, p)} \hfill \\ {\log (E(Y))} \hfill & {Y\sim\,Ga \, (\alpha , \, \lambda )} \hfill \\ {\log (E(Y))} \hfill & {Y\sim\,P \, (\lambda )} \hfill \\ \end{array} } \right.$$
(2)
GAM obtains the most suitable trend line of source data by identifying and accumulating multiple functions. By dealing with the complex nonlinear relationship between the dependent variables and the explanatory variables, the nonparametric regression is fitted. The algorithm iteratively fits and adjusts the function to reduce the prediction error. In fact, GAM pays more attention to nonparametric exploration of data, which is more suitable for analysis and interpretation of the relationship between the response variable and the explanatory variable. The specific analysis of generalized additive model uses R version Rx64.3.4.3 and mgcv package coming from The R-Project for Statistical Computing, i.e., https://​www.​r-project.​org/​.

## 3 Results

### 3.1 Overview of PM2.5 pollution in Beijing

According to the PM2.5 data from Beijing Municipal Environmental Monitoring Center (www.​bjmemc.​com.​cn), it can be calculated that the annual average PM2.5 concentration from December 2016 to November 2017 is 69.46 ug/m3 and the lowest is Yanqing District, northwest of Beijing with an annual average PM2.5 concentration of 52.92 ug/m3 and the highest is Fangshan District, southwest of Beijing with an annual average PM2.5 concentration of 91.13 ug/m3, much higher than 75 ug/m3, the limit of Grade II level stipulated in Ambient Air Quality Standards (GB 3095-2012) with significant spatial gradient. In another word, PM2.5 concentration gradually decreases from south to north (MD = − 19.250, P = 0.004, as shown in Table 2). Besides, 16 districts have experienced the limit of Grade I PM2.5 concentration (55 ug/m3) for over 54% of total days and the limit of Grade II PM2.5 concentration (75 ug/m3) for over 22% of total days.
Table 2
Significance tests of PM2.5 levels for different seasons, regions, daytime and day of week
Variable
Kruskal–Wallis H test
Bonferroni test

Chi-square
P
MD
P
Region#

North versus south
8.661
0.013
− 19.250
0.004
Center versus south

− 8.417
0.258
North versus center

− 10.833
0.074
Season*

Spring versus summer
18.662
0.000
14.109
0.686
Spring versus autumn

− 0.791
1.000
Spring versus winter

− 46.322
0.000
Summer versus autumn

− 14.900
0.578
Summer versus winter

− 60.431
0.000
Autumn versus winter

− 45.513
0.000
Daytime#

7–12 a.m. versus 1 pm–6 pm
45.740
0.000
4.319
0.162
7–12 a.m. versus 7 p.m.–6 a.m.

− 6.455
0.003
1 pm–6 pm versus 7 p.m.–6 a.m.

− 10.775
1.945
Day of week

Weekday versus weekends

Mann–Whitney U test
P = 0.544
* The mean difference is significant at the 0.0083 level for Bonferroni test
# The mean difference is significant at the 0.0167 level for Bonferroni test
During the analysis, it is found that there also exist seasonal fluctuations in the variation of PM2.5 concentration (see Fig. 2). Haze weather occurs 47 times during the research: 9 in spring, 9 in summer, 14 in autumn and 15 in winter, and it usually lasts 1–8 days in spring, 1–2 days in summer, 1–3 days in autumn and 1–9 days in winter featured by large-scale persistent outbreak. Overall, PM2.5 concentration is relatively high in winter and low in summer, and the difference between spring and autumn has no statistical significance (MD = − 0.791, P = 1.000, see Table 2), showing a U-shaped curve all over the year. The main reason is that in winter more coal is burned for heat, increasing the particulate matter discharged to the atmosphere, and from the view of meteorological conditions, pollutant usually diffuses slowly in winter with stable atmospheric stratification of troposphere, no cold air and high-rise buildings in urban areas while the high temperature, exuberant air convection and large precipitation in summer have promoted the deposition of particulate matter. Additionally, Mann–Whitney U test is used to identify the discrepancy of PM2.5 concentration between weekdays and weekends, and the result shows that P = 0.544 (> 0.05), suggesting that the difference of these two groups of data is not significant in statistics.
Expand the above research process and use MATLAB 2014a tool box to conduct autocorrelation analysis on daily PM2.5 concentration to reveal its time-series characteristics, as indicated in Fig. 3. The result has shown that the top and bottom critical values of autocorrelation coefficient are ± 0.163, respectively, and the first-order autocorrelation coefficient is 0.6. It can be seen that PM2.5 concentration of Beijing has strong autocorrelation over time, and it is the most significant from the first to the third order. Besides, PM2.5 concentration also has different diurnal variations in different seasons (as shown in Fig. 4). In comparison with autumn and winter, diurnal variation is stronger in spring and autumn, showing a smooth W shape. Further use Bonferroni correction for detection, and the result has demonstrated that PM2.5 concentration in the morning (7 a.m.–12 a.m.) is far lower than in the nighttime (7 p.m.–6 a.m.) (MD = − 6.455, P = 0.003), but the difference in the daytime (7 a.m.–12 a.m./1–6 p.m.) is not notable. Generally, the lowest daily concentration occurs during the period of 16:00–18:00, while the highest in 9:00–11:00.

### 3.2 Analysis of generalized additive model of PM2.5 and single influence factor

Air pollutants and meteorological factors are the important environmental factors that affect PM2.5 concentration of a place. The air pollutants include SO2 (ug/m3), NO2 (ug/m3), CO (ug/m3) and O3 (ug/m3), and meteorological factors include temperature (°C), atmospheric pressure of sea level (hpa), relative humidity (%), wind speed (km/h), precipitation (mm) and sunshine hours (h). On this basis, adopt generalized additive model to describe the impact a single influence factor and interaction have on PM2.5 concentration with daily average PM2.5 concentration as the response variable and the daily average numeric value of related influence factors as the explanatory variable.

#### 3.2.1 Preanalysis of explanatory variable

According to the analysis result of QQ Chart in SPSS, it can be found that daily average concentration of PM2.5 is approximately subject to gamma distribution (Y ~ (α, λ)). So, use log connection function to connect explanatory variables with response variable subject to gamma distribution via linear combination and calculate the Pearson correlation coefficient between any two explanatory variables to prevent them from distorting the model estimation due to the existence of severe collinearity (|r| > 0.8). It can be known from the calculation result (Table 3) that the correlation coefficient between temperature (t) and sunshine hours (sh) is 0.922 and that between temperature (t) and sea-level atmospheric pressure is − 0.900, suggesting that the longer the same region experiences the sunshine, the higher the temperature is. If other causes such as power are not taken into consideration, the higher the temperature is, the faster the atmosphere expands and rises because of heat, namely that the lower the atmospheric pressure is. Sunshine hour is used to represent temperature index in this paper to avoid concurvity in the construction of multivariable curve model. Moreover, though the correlation coefficient between NO2 and CO has reached 0.856, they are not eliminated as they come from different sources.
Table 3
Pearson correlation coefficient among influencing factors

SO2
NO2
AP
T
WS
SH
RH
P
O3
CO
S02
1
.611**
.354**
− .486**
− .100
− .405**
− .112*
− .110*
− .319**
.539**
NO2
.611**
1
.310**
− .414**
− .312**
− .461**
.176**
− .112*
− .510**
.856**
AP
.354**
.310**
1
− .900**
.013
− .859**
− .202**
− .182**
− .703**
.283**
T
− .486**
− .414**
− .900**
1
− .079
.922**
.225**
.180**
.728**
− .398**
WS
− .100
− .312**
.013
− .079
1
.003
− .592**
− .102
.140**
− .236**
SH
− .405**
− .461**
− .859**
.922**
.003
1
.105*
.230**
.792**
− .422**
RH
− .112*
.176**
− .202**
.225**
− .592**
.105*
1
.452**
− .001
.263**
P
− .110*
− .112*
− .182**
.180**
− .102
.230**
.452**
1
.166**
− .046
O3
− .319**
− .510**
− .703**
.728**
.140**
.792**
− .001
.166**
1
− .389**
CO
0.39**
.856**
.283**
− .398**
− .236**
− .422**
.263**
− .046
− .389**
1
Bold font used to mark the Pearson correlation coefficient $$(|r|)$$ if the absolute value of it is greater than 0.8, which means the two explanatory variables existing severe collinearity
SO2 sulfur dioxide, NO2 nitrogen dioxide, CO carbon monoxide, O3 ozone, AP atmospheric pressure, T temperature, WS wind speed, SH sunshine hours, RH relative humidity
** Correlation is significant at the 0.01 level (2-tailed); * Correlation is significant at the 0.05 level (2-tailed)
It can be known from Fig. 3 that the autocorrelation between PM2.5 concentration in Beijing over time is very strong, which is most obvious in first to third order. But in a short period of time, the development scale, geographical conditions, emissions of industrial pollution and automobile exhaust of a region are relatively fixed, so the change of PM2.5 concentration is mainly related to local meteorological conditions. Next, emphasis is placed on analyzing the correlation between PM2.5 and various meteorological factors located from lag 0 (current PM2.5 concentration) to lag 3 (PM2.5 concentration 3 days ago). With Spearman correlation coefficient (rs) as a measurement criterion, the meteorological elements with a Spearman correlation coefficient bigger than 0.3 (rs > 0.3) are selected as the explanatory variables of the model. And the result is shown as follows.
Integrate the result of Table 4 and take SO2, NO2, CO, O3, wind speed lag 1 (rs = − 0.564, P < 0.01), sea-level atmospheric pressure lag 3 (rs = 0.301, P < 0.01), relative humidity lag 0 (rs = 0.367, P < 0.01) and sunshine hour lag 0 (rs = − 0.308, P < 0.01) as explanatory variables and daily PM2.5 concentration as response variable into the final model (Formula 3). On this basis, spline smooth function is adopted to analyze the impact every explanatory variable has on response variable and the final goodness of fit of the model.
\begin{aligned} \log (E(Y)) = & \,s_{0} + s(SO_{2},bs = cr) + s(NO_{2},bs = cr) + s(CO_{{}},bs = cr) + s(O_{3},bs = cr) \\ & + s(WS lag_{1},bs = cr) + s(AP lag_{3},bs = cr) + s(RH,bs = cr) + s(SH,bs = cr) \\ \end{aligned}
(3)
Table 4
Spearman correlation coefficient among influencing factors
Independent variable
Lag
r s
Sig.(2-tailed)
Wind speed
Lag 0
− 0.356**
0.000
Lag 1
− 0.564**
0.000
Lag 2
− 0.242**
0.000
Lag 3
− 0.075**
0.000
Relative humidity
Lag 0
0.367**
0.000
Lag 1
0.235**
0.000
Lag 2
− 0.028
0.569
Lag 3
− 0.087
0.100
Air pressure
Lag 0
0.073
0.168
Lag 1
0.211**
0.000
Lag 2
0.252**
0.000
Lag 3
0.301**
0.000
Precipitation
Lag 0
− 0.024
0.649
Lag 1
− 0.173**
0.001
Lag 2
− 0.175**
0.001
Lag 3
− 0.154**
0.003
Sunlight hours
Lag 0
− 0.308**
0.000
Lag 1
− 0.252**
0.000
Lag 2
− 0.245**
0.000
Lag 3
− 0.238**
0.000
Bold font used to mark the number if the absolute value of Spearman correlation coefficient $$(|r_{s}|)$$ is greater than 0.3 and the correlation is significant at 0.01 level, otherwise the response variable are not correlated with the explanatory variables
Lag PM2.5 concentration values in different days, rs Spearman correlation coefficient, Sig.(2-tailed) two-tailed tests of significance
** Correlation is significant at the 0.01 level (2-tailed)
The result has shown that all influence factors play a great impact on the change of PM2.5 concentration when P < 0.01, namely that every influence factor is statistically significant as explanatory variable of variation of PM2.5 concentration alone. Among them, the big explanation rates of CO, NO2 and wind speed (WS lag 1) to the variation of PM2.5 concentration (47.5, 44.9, 36.7%) and correction coefficients of determination (0.468, 0.439, 0.357) indicate excellent degree of fitting in the existing model, while the small explanation rate of sea-level atmospheric pressure (AP lag 3) to the variation of PM2.5 concentration (7.56) and the correction coefficient of determination (0.0709) means a bad degree of fitting in the current model (see Table 5).
Table 5
GAM model hypothesis test results between PM2.5 concentration and single influencing factors
Smooth term
df
Ref.df
F
P
Deviance explained (%)
SO2
2.159
2.698
43.18
< 2E−16***
24.7
0.243
NO2
6.376
7.632
38.18
< 2E−16***
44.9
0.439
CO
4.563
5.578
57.57
< 2E−16***
47.5
0.468
O3
7.233
8.252
15.85
< 2E−16***
24.7
0.259
AP lag 3
1.798
2.271
12.36
2.73E−06***
7.56
0.0709
WS lag 1
5.221
6.275
32.24
< 2E−16***
36.7
0.357
RH
2.391
3.007
23.75
3.26E−14***
17.0
0.165
SH
7.569
8.499
7.447
1.93E−09***
15.7
0.139
*, **, *** Significant level at 5, 1, 0.1%
Besides, when degree of freedom (df) is 1, the function is a linear equation, suggesting that there exists a linear relation between explanatory variables and response variable. When it is bigger than 1, the function is a nonlinear curve equation and the bigger it is, the more significant the nonlinear relation is. Among eight explanatory variables in this experiment, there is certain nonlinear relationship between SO2, relative humidity (RH) and sea-level atmospheric pressure (AP lag 3) and PM2.5 (with a degree of freedom of 2) and significant linear relationship between other factors and PM2.5. So, variation of PM2.5 concentration is a sophisticated nonlinear time variation series driven and affected by multiple factors. By building generalized additive model between explanatory variables and PM2.5 concentration, the effect plot between explanatory variables and PM2.5 concentration can be obtained. In Fig. 5, the independent influence every predictive variable has on PM2.5 is depicted: the dashed line refers to the point-to-point standard deviation of fitting additive function, namely the top and bottom limitations of confidential interval and the full line is the smooth fitting curve of PM2.5 concentration. X-axis is the measured value of various explanatory variables, and Y-axis is the smooth fitted value of different explanatory variables over PM2.5 concentration.
It can be found from Fig. 5 that the variation of PM2.5 concentration is affected by multiple influence factors. Interact the explanatory variables and analyze the impact the interaction terms have on variation of PM2.5 concentration. It can be known from the calculation result of Table 6 that the interaction terms have very big numeric value in degree of freedom, suggesting that they have significant nonlinear relations with variation of PM2.5 concentration. Additionally, 28 interaction terms in the model equation have got through the significance test with an explanation rate between 19.6 and 58.0% and a significant rise compared with single-factor model, proving that the fitting degree of the model has reached a high standard. The top five places in fitting degree are CO-WS lag 1 (58%), CO-SH (56.2%), SO2-WS lag 1 (54.8%), NO2-WS lag 1 (51.9%) and CO–RH (50.7%), and they are combinations of precursors and meteorological elements. It states that PM2.5 concentration changes mainly under the interaction between air pollutants and meteorological elements from one side, and the calculation result has also shown that wind speed (WS lag 1) has played a decisive role in the entire diffusion, and it can largely explain the entire change of PM2.5 concentration.
Table 6
GAM model hypothesis test results between PM2.5 concentration and interaction of influencing factors
Cross term
SO2-NO2
SO2-CO
SO2-O3
SO2-APlag 3
SO2-WSlag 1
SO2-RH
SO2-SH
df
8.41
17.63
12.20
4.60
15.40
8.59
9.26
DE
45.9%
49.3%
38.4%
25.1%
54.8%
45.1%
27.5%
0.446
0.467
0.362
0.242
0.527
0.437
0.256
Cross term
NO2-CO
NO2-O3
NO2-RH
NO2-APlag 3
NO2-WS lag 1
NO2-SH
CO-O3
df
9.127
16.56
3.241
9.716
2.003
10.59
24.55
DE
50.0%
50.3%
48.9%
44.7%
51.9%
46.0%
48.8%
0.487
0.479
0.484
0.431
0.516
0.444
0.451
Cross term
CO-APlag 3
CO-WSlag 1
CO–RH
CO-SH
O3-APlag 3
O3-WSlag 1
O3-RH
df
21.42
18.10
19.78
18.76
15.46
13.98
10.88
DE
48.1%
58.0%
50.7%
56.2%
34.0%
45.1%
31.4%
0.448
0.558
0.479
0.538
0.311
0.429
0.293
Cross term
O3-SH
APlag 3-WSlag 1
APlag 3-RH
APlag 3-SH
WSlag 1-RH
WSlag 1-SH
RH-SH
df
12.51
11.54
8.975
18.64
16.27
21.26
22.17
DE
29.3%
42.8%
32.6%
19.6%
39.3%
48.6%
37.8%
0.268
0.409
0.309
0.152
0.364
0.453
0.338
DE deviance explained
*, **, *** Significant level at 5, 1, 0.1% (all terms: P < 2E−16***)
Take average wind speed (WS lag 1) as the key factor and analyze the impact the interaction between the average wind speed and concentration of air pollutants has on PM2.5 concentration as shown in the Fig. 6.
When CO concentration is low, PM2.5 concentration slowly decreases with the increase in average wind speed (WS lag 1), and when the average wind speed (WS lag 1) is quite low, the concentration of PM2.5 sees a fluctuant increase with the increase in CO concentration. It means that wind speed can dilute CO concentration in the air and it significantly affects the mixing effect of pollutants when the wind speed is low and increases PM2.5 concentration.
NO2 and SO2 have similar variation relations with average wind speed (WS lag 1). When NO2/SO2 is low in concentration, average wind speed can diffuse and dilute and stabilize PM2.5 concentration within a relatively low interval. However, with the increase in NO2/SO2 concentration, average wind speed plays a less and less significant role, indicating that their concentration in the air has a critical value. Once it exceeds that value, the average wind speed plays a more significant role on the mixture of NO2 and SO2 and promotes the occurrence of secondary chemical reaction, so the pollutants in the air cannot be effectively diluted.
As O3 is mainly distributed in the stratosphere with a height of 10–50 km, average wind speed (WS lag 1) barely has any influence. It can be learnt from Table 3 that there is certain negative correlation between O3 and NO2 and the increase in NOx will increase secondary nitrate particulate matter and produce release PM2.5. On the other hand, the low visibility caused by rise of PM2.5 concentration will suppress the production of O3. So, it can be concluded that there also exists negative correlation between PM2.5 concentration and O3 concentration. In recent years, PM2.5 concentration in China has dropped a bit, but O3 pollution has become increasingly serious, especially in summer. It seems that O3 may become the primary pollutant in place of PM2.5. Therefore, PM2.5 and O3 shall be taken into collaborative control in the governance.
In the previous research on the relationship between PM2.5 and meteorological elements, the majority focuses on wind speed and proves that there is negative correlation between them [26, 27], which has also been verified in this paper. What is different is that based on the precursors of PM2.5, time autocorrelation has also been taken into consideration and time-delay characteristic of meteorological elements have also been taken into test. The calculation result has shown that SO2, NO2, CO, O3, wind speed lag 1 (rs = − 0.564, P < 0.01), sea-level atmospheric pressure lag 3 (rs = 0.301, P < 0.01), relative humidity lag 0 (rs = 0.367, P < 0.01) and sunshine hour lag 0 (rs = − 0.308, P < 0.01) are the variables most correlated with daily average PM2.5.

## 4 Conclusions

The pollution of fog and haze, as one of the most severe environmental pollution problems in recent years in China, has a great effect on the physical and psychological health and regular travel of the public and the economic development of the state. In the critical period of economic development in China, it is urgent to solve fog and haze pollution. Generalized additive model is adopted in this paper to analyze the spatial–temporal characteristics of pollution of fog and haze and the impact brought by various influence factors and the interactive terms in Beijing and the surrounding regions, and the entire diffusion process is analyzed on this basis. The key conclusions are as follows.
(1)
There exists a significant spatial gradient in pollution of fog and haze in Beijing and the surrounding areas, as indicated by progressive decrease from south to north (MD = − 19.250, P = 0.004, as shown in Table 2), which is mainly the result of mutual accumulation of regional pollution. The northern area is surrounded by mountains and plenty of green vegetation can purify the air to a certain degree [16]. On the other hand, as the southern area is adjacent to such severely polluted regions as Tianjin Municipality and Hebei Province, external influence is more significant [1719]. So, regional collaborative prevention and treatment is required to control the pollution of PM2.5 in Beijing and special attention shall be paid to the transmission of pollutants from the southwest and the control of major sources.

(2)
PM2.5 concentration fluctuates greatly in different reasons. As a whole, the concentration of PM2.5 is relatively high in winter and low in summer with a U-shaped distribution all year round. According to “Measures of Beijing Municipality for Administration of Heat Supply and Use,” the heating period of this Municipality is from November 15 of the current year to March 15 of the next year, so PM2.5 concentration goes up in winter mainly because household heating increases and causes more particulate matter discharged into the air [17, 18, 20]. Earlier, the seasonal characteristics of spring drought, few rains and strong wind as well as unreasonable human activities in Beijing have frequently resulted in large-scale outbreak of sand storm, but in the experiment result of this paper, PM2.5 concentration does not increase greatly in spring, suggesting that the sandstorm source control project in Beijing–Tianjin–Hebei Region has worked well. Besides, some research achievements have once pointed out that there is a difference in pollutant concentration between weekdays and weekends [21, 22], but according to the experiment result of Mann–Whitney U test in this paper, the foregoing difference is not statistically significant and the credit should be given to the implementation of vehicle ban policy in China.

(3)
The daily change rule of PM2.5 is subject to the height of boundary layer [23]. According to the calculation result, the lowest concentration usually occurs during 16:00–18:00 and the highest during 9:00–11:00 every day. Generally speaking, the peak PM2.5 concentration in the morning is caused by human activities, while the fall of PM2.5 concentration in the afternoon is because the increase in height of boundary layer has accelerated the diffusion of PM2.5. In the nighttime, the height of boundary layer drops again and human activities increase, so PM2.5 concentration rises again. Therefore, PM2.5 concentration every morning (7–12 a.m.) is much lower than every night (7 p.m.–6 a.m.) (MD = − 6.455, P = 0.003). Especially in winter, heating increases from coal burning and so does the particulate matter released to the air. Besides, less solar radiation also moves the time when the height of boundary layer falls in advance, so the nighttime PM2.5 level is relatively high in winter [24, 25].

(4)
PM2.5 concentration changes mainly due to the interaction between air pollutants and meteorological elements [26, 27]. In the entire diffusion process, average wind speed (WS lag 1) plays a decisive role and it affects the diffusion of CO and the occurrence of secondary reaction. For NO2 and SO2, average wind speed (WS lag 1) has a critical value. The concentration of NO2 and SO2 remains stable below the critical value, and secondary reaction occurs when exceeding the critical value. For O3, as it is mainly distributed in the stratosphere with a height of 10–50 km, average wind speed (WS lag 1) has no effect on it, but it can be concluded that there is negative correlation between PM2.5 concentration and O3 concentration through the calculation result. Therefore, in the governance, they shall be taken into collaborative control.

## Acknowledgements

This work was supported by the National Social Science Foundation of China (No.16BGL145).

## Compliance with ethical standards

### Conflict of interest

We declare that we have no financial and personal relationships with other people or organizations that can inappropriately influence our work; there is no professional or other personal interest of any nature or kind in any product, service and/or company that could be construed as influencing the position presented in, or the review of, the manuscript entitled.

## Unsere Produktempfehlungen

### Premium-Abo der Gesellschaft für Informatik

Sie erhalten uneingeschränkten Vollzugriff auf alle acht Fachgebiete von Springer Professional und damit auf über 45.000 Fachbücher und ca. 300 Fachzeitschriften.

### Springer Professional "Wirtschaft"

Online-Abonnement

Mit Springer Professional "Wirtschaft" erhalten Sie Zugriff auf:

• über 58.000 Bücher
• über 300 Zeitschriften

aus folgenden Fachgebieten:

• Bauwesen + Immobilien
• Finance + Banking
• Management + Führung
• Marketing + Vertrieb
• Versicherung + Risiko

Testen Sie jetzt 30 Tage kostenlos.

### Springer Professional "Technik"

Online-Abonnement

Mit Springer Professional "Wirtschaft" erhalten Sie Zugriff auf:

• über 50.000 Bücher
• über 380 Zeitschriften

aus folgenden Fachgebieten:

• Automobil + Motoren
• Bauwesen + Immobilien
• Elektrotechnik + Elektronik
• Energie + Umwelt
• Maschinenbau + Werkstoffe

Testen Sie jetzt 30 Tage kostenlos.

Weitere Produktempfehlungen anzeigen
Literatur
Über diesen Artikel

### Weitere Artikel der Ausgabe 5/2019

Zur Ausgabe

S.I. : Emerging Intelligent Algorithms for Edge-of-Things Computing