Variation characteristics and prediction of pollutant concentration during winter in Lanzhou New District,China

DongYu Jia,XiaoXia Li,XiaoQing Gao,LiWei Yang

1.College of Geography and Environmental Engineering,Lanzhou City University,Lanzhou,Gansu 730070,China

2. Northwest Regional Climate Center/Key Laboratory of Arid Climatic Change and Reducing Disaster of Gansu Province/Key Open Laboratory of Arid Climatic Change and Disaster of China Meteorological Administration/Institute of Arid Meteorology,Lanzhou,Gansu 730020,China

3. Key Laboratory of Land Surface Process and Climate Change in Cold and Arid Regions/Northwest Institute of Eco-Environment and Resources,Chinese Academy of Sciences,Lanzhou,Gansu 730000,China

ABSTRACT PM2.5 and PM10 were the main air pollutants during winter in Lanzhou New District,China.In this paper,WRF model output combined with hourly monitoring data of pollutant concentration was used to analyze characteristics of the concentration change and to study the relationship between meteorological elements and PM10/PM2.5 in Lanzhou New District in January, 2018. Meanwhile, the concentration changes of PM2.5 and PM10 were predicted by wavelet analysis combined with BP neural network.The results show that:(1)Due to the cold front process in winter,PM2.5 was negatively correlated with the water vapor mixing rate. PM10 was positively correlated with air temperature and negatively correlated with air pressure. (2)There was an inversion layer in the atmosphere near the high value day of PM2.5 and PM10, the surface was controlled by low pressure, low wind speed, and the situation of low value day of PM2.5 was the opposite. On the day of high value of PM10, the air temperature below 600 hPa was higher, and the wind speed near the surface was also higher. (3)Wavelet analysis combined with BP (Back Propagation) neural network had a good prediction effect on PM2.5, which could basically reflect the hourly change of PM2.5 concentration.However,the simulation effect of PM10 was poor,and the input parameters of surrounding pollutants should be added to improve the prediction effect.

Keywords:PM2.5;PM10;WRF;Wavelet neural network;Lanzhou New District

1 Introduction

With the acceleration of urbanization, the problem of environmental pollution becomes increasingly prominent, and air quality has a great impact on people's production and life(Dinget al.,2009;Luoet al.,2019). PM10(particulate matter with aerodynamic diameters less than 10 μm)and PM2.5(particulate matter with aerodynamic diameters less than 2.5 μm)have attracted close attention from the government and the public in recent years due to their ability to affect air visibility and human cardiopulmonary function (Wuet al., 2019). Studies have shown that the pollution of PM2.5in the atmosphere in northern China is significantly heavier in winter than in summer, especially in the winter heating season(Zhanget al.,2006).

There are many influential factors for air pollution, which are mainly divided into two aspects: one is that local air pollutant emissions reach a certain concentration; the other is that local meteorological circulation is conducive to the accumulation of pollutants in the atmosphere (Zhanget al., 2014; Karenet al., 2020; Zhaoet al., 2020). In the case that pollutant emission concentration is not easily changed, atmospheric circulation and meteorological factors become the main factors affecting pollutant concentration. Researchers have found that wind direction, downwelling solar radiation and around terrain are important factors for the transport and spatial-temporal distribution of PM2.5and PM10(Yanget al., 2016; Zhenget al., 2017; Zhanget al., 2019; Fanet al., 2020).The research of Muet al. (2011) and Panet al. (2020)shows that boundary layer height and boundary layer structure had a great impact on the concentration of PM10and PM2.5pollutants. Among them, the concentration of PM2.5in the stable boundary layer structure is significantly higher than that in the convective boundary layer structure. Panet al. (2019) found that the diurnal variation of boundary layer height had a strong influence on the daily cycle of PM2.5concentration. Among them, the daily cycle of PM2.5in winter is bimodal, which corresponds well to low boundary layer conditions. Zhanget al. (2019) found that the differences of PM2.5among five cities are most likely associated with emission sources,meteorology and topography. Bimodal distribution of PM2.5should be associated with both human activity and planetary boundary layer (PBL) variation. Zhanget al. (2006)found that the relationship between meteorological factors and PM2.5pollutant concentration varies according to different cities, among which relative humidity, wind speed and temperature are the main variables.Zhaoet al.(2013)found that precipitation,relative humidity and wind speed are closely related to the change of pollutant concentration in Lanzhou.Moreover, the influence of precipitation on air pollutant concentration was also found in Alaska (Timothyet al.,2010)and Beijing(Sunet al.,2019).

Lanzhou New District is located at the intersection of the Mongolian Plateau, the Loess Plateau and the Qinghai-Tibetan Plateau, and its wind speed is greatly affected by the terrain (Liet al., 2017).Therefore, meteorological conditions have a greater impact on the diffusion of pollutants in the region. At the same time, many large chemical enterprises have moved to Lanzhou New District. Large releases of precursor gases for secondary formation of fine aerosol particles, along with the fast growth of fine particles (Zhaoet al., 2018), can have a serious impact on the environment of the study area. In recent years,there are some studies on the pollution of Lanzhou area. For example, Zhaoet al. (2020) found that the PM2.5concentration in Lanzhou area was mainly affected by automobile exhaust, coal burning and secondary inorganic pollution sources, and significantly increased in dust weather. The research of Liet al.(2017) found that the prevailing wind direction of each layer height in Duanjiachuan area of Lanzhou New District was the same, which was northeast wind.However,the prevailing wind direction and pollution coefficient at 10 m were significantly different from those at 70 m. Zhaoet al. (2003) established a winter pollution prediction equation for Lanzhou based on the visibility and circulation field of Lanzhou,which could basically predict the change of pollutant concentration by using meteorological factors.However, due to the limitation of observational data,there is limited research on the change, prediction of pollutants and their relationship with meteorological elements in Lanzhou New District.

As the fifth new district of China, Lanzhou New District plays an important role in the production and life of northwest China. At the same time, during the construction of the new district in the past decade, the population and the underlying surface of the area have undergone great changes (Liuet al., 2016).Therefore, analysis of the characteristics of air pollution changes in this region is conducive to the establishment of a relatively complete air quality prediction mechanism in this region in the future (Zhanget al., 2010). This study aims to explore the relationship between different meteorological factors and pollutant concentration by using WRF model combined with meteorological and pollution observation data in Lanzhou New District. Then, meteorological factors were taken as input terms to predict pollutant concentration by using wavelet analysis combined BP neural network.

2 Data and methods

2.1 Data collection

Meteorological elements observation data are from three 70 m wind towers (Observation point 1,Observation point 2, Observation point 3) established by Gansu Meteorological Bureau, including temperature, air pressure, horizontal wind speed and direction meteorological observation data at different heights(10, 30, 50, 70 m) every half hour interval, from 1 January to 31 January, 2018. The pollutant data are based on hourly observation data of PM2.5and PM10from two continuous observation points (Observation point 4 and Observation point 5) provided by the Environmental Protection Bureau of Gansu Province, from January 1 to January 31, 2018. Among them, the instrument type of pollutant concentration detection (including PM10and PM2.5) at observation point 4 is XHPM2000E, the analysis method isβray analysis, and the detection standard of PM10is HJ654-2013, the standard of PM2.5is HJ653-2013.The instrument type of pollutant concentration detection (including PM10and PM2.5) at observation point 5 is BAM-1020, the analysis method is β ray analysis,and the detection standard of PM10is HJ654-2013,the standard of PM2.5is HJ653-2013. In the observation simulation of Lanzhou New District, the geographical location and elevation of each observation point is presented in Table 1. The average altitude of the observation point is more than 2,000 m above sea level, and the terrain of the observation area is flat.The measured data can basically represent the boundary layer characteristics of Lanzhou New District and surrounding areas.

Table 1 Geographical location of observation point

2.2 Model set-up

In this study, the mesoscale weather model WRF v3.9.1 was used to simulate the boundary layer characteristics of Lanzhou New District.The mode adopts three-layer nesting. The number of cells nested from the outer layer to the inner layer is 88×70, 109×91,34×52, respectively. The resolutions are 9 km, 3 km,and 1 km,respectively(Figure 1).The initial field and lateral boundary conditions were obtained from NCEP (National Centers for Environmental Prediction) GDAS (Gwent Drug & Alcohol Service) Final 0.25°×0.25° analysis data, and the data were read and updated once every 6 h.The simulation starts at 00:00 on 30 December, 2017 and ends at 23:00 on January 31, 2018. In this study, simulated data from 00:00 on December 30 to 23:00 on December 31 were taken as spin-up time. The parameter setting in WRF is presented in Table 2.

Based on the data of MODIS30 (Moderate-resolution Imaging Spectroradiometer, MODIS), the Landsat30 (Launch of the Earth Resources Technology Satellite) and the new district planning map,the artificial surface area has been revised. Errors in modeled winds are estimated by comparing WRF predictions with profiler measurements of the wind components at a height of 70, 50, 30 and 10 m.Using data from January, 2018, the root mean square (RMS) errors in horizontal wind speed at all height is 2.28 and the RMS errors in wind direction at all height is 92.1. According to the research of Zhaoet al. (2013), some of this difference can be attributed to the fact that profiler winds are measured at a single site while the WRF winds are averages over one square kilometer area.After adding the artificial undersurface in Lanzhou New District, the simulated wind speed error is significantly reduced compared with that before correction, and the RMS error decreases by 0.2. At the same time, it is found that the simulation effect of wind speed is good, and its correlation coefficient with the observed value reaches 0.7.

Figure 1 Three-layer nested domain of WRF(The black points are the observation station locations)

2.3 Wavelet analysis combined with BP neural network

Figure 2 is a flowchart of the wavelet neural network to predict pollutant concentration. In this study,by using three-layer wavelet decomposition for pollutant concentration, the approximation signal A3 and the detailed signals D3, D2 and D1 are obtained. In this study, meteorological variables closely related to pollutant concentration (wind speed, air temperature,pressure, precipitation, water vapor mixing rate and boundary layer height) were selected as input parameters of training samples. The training samples were observations from January 1 to January 28, 2018, and the test samples were observations from January 29 to 31, 2018. In this paper, the number of hidden nodes selected for the neural network model is 7. Matlab 2018a was used to model the prediction program, and the predicted values were reconstructed by wavelet to obtain the predicted values of pollutant concentration.Wavelet transform and wavelet neural network are described in detail below.

Table 2 The parameter setting in WRF

Figure 2 The forecasting process of pollutant concentration by wavelet transform and neural network

1) Wavelet transformation

Wavelet transformation is to decompose the mother wavelet into a series of sub-signal sequences through displacement and scaling on the time axis.Because wavelet transform can reflect local data information which cannot be obtained by Fourier analysis,it has a better ability to filter signals. Wavelets are classified as real or complex analytic wavelets. Real wavelet is often used to detect and filter sharp signals.As it can separate amplitude and phase components,complex analytic wavelet is often used to measure transient and time evolution of frequency. One of the most widely used parent wavelets in wavelet analysis is Morlet wavelets. The wavelet (Malletet al., 1999)is expressed as follows:

where'c'is the frequency of the mother wavelet and'i'is imaginary.Its Fourier transform is as follows:

where 'c'is the frequency of the mother wavelet,'w'is the variate.

For continuous wavelet of any function, the expression of wavelet decomposition algorithm is as follows(Huang,2004):

where,'n'is the initial data sequence off(t);j-number of decomposition layers;Aj-f(t)is the wavelet coefficient of the function approximates the signal at the first layer;Dj-f(t) is the wavelet coefficient of the detail signal at the first layer.

2) BP neural network

In this paper, the BP wavelet neural network is a three-layer network structure,mainly including the input layer, the hidden layer and the output layer. The input layer and the hidden layer are connected by a weight matrix, as are the hidden layer and the output layer. The structure of the wavelet neural network used in this paper is presented in Figure 2.The activa-tion function of the hidden layer is Morlet (Malletet al., 1999; Wong and Qi, 2009) wavelet function, as shown in Equation(3).

where't'is the variate.

The activation functionf(x) of the output layer is the log-sigmoid function, and the form off(x) is shown in Equation(4):

3 Results

3.1 Analysis of variation characteristics of pollutant concentration in Lanzhou New District

Zhaoet al. (2015) found that air pollution in Lanzhou was mainly affected by two types of weather in winter. Among them, static and stable weather was mainly caused by high environmental stability and low wind speed,which was not conducive to the diffusion of pollutants.

Combined with the temperature change and weather field in the study area, it can be seen that there were several cold fronts passing through during the study period.The maximum mixing layer thickness in Lanzhou was the smallest before the cold front,so the particulate pollution was mainly affected by the cold front. Referring to Figure 4, it can be seen that the increase of pollutant concentration was closely related to the increased transport associated with higher wind speeds (during cold front transit), and at the same time,the pollutant concentration will also decrease after the frontal transit due to the wet scavenging of rain and snow.

Figure 3 The temperature change of Lanzhou New District in January,2018

As can be seen from Figure 4, in January, 2018,the daily average peak concentration of PM10at observation point 4 and observation point 5 were 172 μg/m3,124 μg/m3, and the valley values were 74 μg/m3,43 μg/m3, respectively. The daily average peak concentration of PM2.5at observation point 4 and observation point 5 were 95 μg/m3and 84 μg/m3, and the valley values were 27 μg/m3and 19 μg/m3, respectively.The study found that PM10in Lanzhou New District was easily affected by local factors, and the overall value of the two observation points was different, but the trend of change was the same. However, due to small population density in the study area, there was little difference in the value of PM2.5between the two observation stations.

Figure 4 Daily average concentration change of PM10 and PM2.5 in January,2018

China promulgated the ambient air quality standard for PM10in 1996 (State Environmental Protection Bureau, 1996), and increased the limit for PM2.5in 2012. Among them, the concentration limits of PM10in the first and second levels were 50 μg/m3and 150 μg/m3, respectively, and the concentration limits of PM2.5were 35 μg/m3and 75 μg/m3, respectively(State Environmental Protection Administration, 2012).Combined with Figure 5, it can be seen that the number of days when the observation point 4 was at the second-order concentration limit was higher than observation point 5. Because there was a main road near observation point 4, and the traffic density was obviously higher than that of Zhouqu Middle School.According to the ambient air quality standard, the firstgrade compliance rates of PM10concentration in observation point 4 and observation point 5 were 0%and 3.23%, respectively, and the second-grade compliance rates were 83.87% and 100%, respectively. In observation point 4 and observation point 5,the first-level PM2.5compliance rate was 19.35%and 29.03%,respectively, and the second-level compliance rate was 83.87% and 93.54%, respectively. The pollutant concentrations at observation points 3 and 4 were mainly concentrated within the second-level limit, which was considered as moderate pollution,in January 2018.

Figure 5 Days distribution of PM10(a)and PM2.5(b)at different concentration limits in January,2018

3.2 The relationship between local meteorological conditions and pollutant concentration

Meteorological conditions were conducive to the diffusion and dilution of local pollutants,and it can also affect the secondary formation(gas-particle conversion), which are closely related to air quality (Liuet al., 2014; Zhaoet al., 2018). Therefore, this study used Pearson correlation coefficient to analyze the correlation between PM10/PM2.5concentrations and different meteorological factors.Among them,the correlation coefficient in Table 3 is calculated by using hourly meteorological observation data and pollution concentration data in Lanzhou New District. The results show that in winter, due to the low surface temperature the increase in temperature may lead to the formation of temperature inversion layers, which are unfavorable conditions for atmospheric movement and can cause the accumulation of PM2.5and PM10.Humidity exerts a positive influence on PM2.5and PM10concentrations through more vapor and notably increases the mass concentration, which is the hygroscopic increase and accumulation of PM2.5and PM10.Atmospheric pressure negatively influences PM2.5and PM10concentrations through the formation of lower-level wind convergence, which affects the accumulation and diffusion of PM2.5and PM10. Meanwhile, atmospheric pressure can negatively influence PM2.5and PM10concentrations by affecting other factors like wind speed. These results are consistent with Chen's results (Chenet al., 2020).According to the research conclusion of Wuet al. (2019), both PM10and PM2.5were not highly correlated with precipitation in the study period, which should be caused by the absence of significant precipitation in the study period.

Table 3 Correlation between pollutant concentration and different meteorological elements

According to the research of Zhanget al. (2014),meteorological conditions such as temperature inversion layer, stationary weather and weak wind all affect the concentration of PM2.5. In order to study the relationship between pollutant concentration and meteorological elements in Lanzhou New District, this paper selects the day when PM2.5and PM10concentrations were at peak and valley values respectively, analyzes the diurnal variation characteristics of pollutant concentration,and analyzes the relationship between meteorological elements and pollutant concentration based on the weather situation at 14:00, in January, 2018. In comparison with Du's study (Duet al., 2001), WRF can accurately simulate the boundary layer height variation in Lanzhou region.Therefore, the boundary layer height of this study adopts WRF output data.

According to Figure 6, when the air quality was light pollution(e.g.,Figure 6b and Figure 6d),the pollutant concentration around 00:00 and 20:00 (Beijing Time)was relatively stable due to the influence of stable atmospheric stratification. However, when heavy pollution occurs, the atmospheric boundary layer height before sunrise was low, which was not conducive to the diffusion of pollutants, so the pollutant concentration peak appeared around 00:00 (Beijing Time).Among them,the conclusions on the change of pollutant concentration during sunrise and sunset were consistent with those of Yanget al. (2020), Baiet al. (2018) and Guiet al. (2019).After sunrise, stable stratification was damaged and the boundary layer was elevated. However, by comparing Figure 6 and Figure 7, it can be seen that the increase of boundary layer height at noon had little impact on PM2.5concentration (no matter in light or heavy pollution days).Therefore, based on the peak value of pollutant concentration change,it could be seen that PM2.5pollutant concentration change in the study area was mainly affected by pollution sources.

Figure 6 Concentration change of PM2.5 and boundary layer height on heavy(a,January 5;c,January 28)and light pollution days(b,January 5;d,January 28)

According to Figure 6 and Figure 7, when PM2.5concentration was high, the corresponding surface temperature was high, and the study area was mainly affected by warm air mass during this period, while the ground was controlled by low pressure. Combined with the wind speed and temperature profile of observation point 4 and observation point 5, it can be seen that on days with high PM2.5concentration, an obvious inversion layer appeared in the air temperature from 850 hPa to 650 hPa, which was closely related to the influence of warm advection. Especially at night, the double inversion layer appears at the height of 700 hPa and 500 hPa, which was very unfavorable to the diffusion of pollutants. At the same time, according to the wind profiles of the two observatories in Figure 7, wind speed near the surface, with high PM2.5values, was relatively low. Combined with the above factors, the concentration of PM2.5was high.During days with low PM2.5concentration, the study area was controlled by cold and high pressure,and the local area presented a downdraft. Moreover, wind speed near the ground was relatively high, which was conducive to the diffusion of pollutants. In general,high wind speed was conducive to the diffusion of PM2.5pollutants, but may cause an increase in PM10concentration.

Figure 7 Weather characteristics at 14:00 on the day with the highest concentration(a,b,c,January 5)and the lowest concentration(d,e,f,January 28)of PM2.5(Figures a,d surface temperature and wind speed,Figures b,e near formation water vapor mixing rate,Figures c,f wind speed and temperature profile wind profile:black,temperature profile:blue)

Due to the low vegetation coverage in winter and spring, dust transport was the main factor causing the increase of PM10concentration. As seen in Figure 9a,Figure 9b,in the sunrise,sunset and noon,the concentration of PM10were significantly increased. Combined with Figure 6 and Figure 7,it shows that the influence of the boundary layer height of PM10in the study area was more significant than PM2.5. At noon,the turbulence caused by solar heating effect increases significantly, which would increase the vertical movement of pollutants near the surface. At the same time, increased anthropogenic activities in the region (e.g., transportation) also contributed to increased PM10as a result of dust. The inversion structure at the top of the boundary layer has an obvious inhibitory effect on the vertical transport of pollutants,while the higher top height of the boundary layer in the afternoon make the atmospheric boundary layer have a large volume capacity of pollutants,thus effectively reducing the PM10concentration in that period.As can be seen from Figure 8,on the day with high PM10concentration, the wind around the study area presented a small local circulation. Combined with the wind temperature profile, it can be seen that the high wind speed from the near surface to the height of 700 hPa was the reason for the heavy dust weather.

3.3 Prediction of pollutant concentration change in Lanzhou New District

According to the above analysis, meteorological elements were closely related to PM2.5and PM10concentrations.Thus, this article selected air temperature,water vapor mixture rate, surface temperature, wind speed, precipitation and air pressure boundary layer height, the hourly data from January 1 to 28 as input parameters, PM2.5and PM10concentration respectively as the output parameters by wavelet analysis combined with BP neural network training, took the data from January 29 to 31 as test data, the overall simulation results as presented in the figure below.

Figure 8 Weather characteristics at 14:00 on the day with the highest concentration(a,b,c,January 13)and the lowest concentration(d,e,f,January 20)of PM10(Figures a,d surface temperature and wind speed,Figures b,e near formation water vapor mixing rate,Figures c,f wind speed and temperature profile wind profile:black,temperature profile:blue)

As can be seen from Figure 10, overall wavelet analysis combined with BP neural network could predict the overall trend of PM2.5and PM10pollutant concentrations. Among them, according to the above analysis, due to the concentration of PM2.5mainly caused by human activities, our study area was not affected by pollutants transport, so the overall simulation effect was good, but slightly higher than the observation data of simulation data before and after 13:00 on January 30, this should be due to overestimate the boundary layer height of PM2.5caused by dry fall after sunrise. Since PM10was affected by both foreign and local pollutants, it was not possible to accurately predict the concentration change of PM10by only using local meteorological elements.The surrounding meteorological and environmental elements should be added as input parameters in subsequent studies, so as to further improve the accuracy of prediction.

4 Conclusions and discussions

In this paper,WRF model was used to analyze the relationship between pollutant concentration and meteorological elements in Lanzhou New District, Gansu Province in January 2018. In addition, the wavelet neural network method was used to predict the change of pollutant concentration, and the uncertainty of the model for visibility prediction method was discussed. This study found that after updating the data of the underlying surface and artificial surface area,WRF simulation effect was improved and the error of basic meteorological elements simulation was reduced,which could meet the research needs.

Figure 9 Concentration change of PM10 and boundary layer height on heavy(a,January 13;b,January 20)and light pollution days(c,January 13;d,January 20)

Figure 10 Prediction renderings of PM2.5(a)and PM10(b)at different observation points

Meteorological conditions were conducive to the diffusion and dispersion of local pollutants, and it can also affect secondary formation (gas-particle conversion), which are closely related to air quality.Among them, surface temperature, wind speed and boundary layer height would affect the transmission of pollutants. The water vapor mixing rate and precipitation would reduce the change of pollutant concentration through wet sedimentation. In addition, weather processes, such as the passage of a cold front, also influenced changes in pollutant concentrations. Both wind speed and inversion layer had great influence on pollutant concentration in the study area. Among them,the appearance of inversion layer was not conducive to the diffusion of PM2.5and PM10pollutants, thus maintaining a large concentration at night. High wind speed was conducive to the diffusion of PM2.5pollutants, but it might increase the concentration of PM10due to floating dust.

Wavelet analysis combined with BP neural network had a better prediction effect on PM2.5than PM10, because the PM2.5concentration in Lanzhou New District mainly came from local man-made activities.At the same time, due to the low wind speed in winter,the relationship between PM10and meteorological variables was not significant.Therefore,the simulation effect was affected.

It should be pointed out that due to the limitation of observational data, this paper only analyzed the WRF model simulation data and pollutant observation data in January 2018, and did not verify with historical events.At the same time, no significant precipitation occurred during the study period, which did not show the wet precipitation effect of precipitation on air pollution.In the future,pollution changes in different seasons and the corresponding circulation characteristics in Lanzhou New district should be analyzed by combining observational data and numerical simulation data, so as to obtain a general conclusion. At the same time, PM10dominated the pollution characteristics in this area deserve attention. In addition, the route and source of pollution dispersion should be further considered, which could provide some references for pollution prevention and control in Lanzhou New District in the future.

Acknowledgments:

This study was supported by research and development plan of Gansu Province in 2018 "Experimental study on atmospheric environment characteristics of near-ground boundary layer in Lanzhou New District serving fine functional zoning" (18YF1FA100), the Opening Fund of Key Laboratory of Land Surface Process and Climate Change in Cold and Arid Regions, CAS (Grant No. LPCC2018006) and the Lanzhou City University Doctoral Research Initiation Fund (Grant No. LZCU-BS2019-13). The authors are grateful to the anonymous reviewers for their constructive comments.