The Impacts of Different PBL Schemes on the Simulation of PM 2 . 5 during Severe Haze Episodes in the Jing-JinJi Region and Its Surroundings in China

1Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters/Key Laboratory for Aerosol-Cloud-Precipitation of China Meteorological Administration, Nanjing University of Information Science and Technology, Nanjing 210044, China 2State Key Laboratory of SevereWeather/Institute of Atmospheric Composition, Chinese Academy ofMeteorological Sciences (CAMS), CMA, Beijing 100081, China 3Yantai Meteorological Bureau, Yantai 264003, China


Introduction
Owing to its population explosion, accelerated urbanization, and globalization, China-the country with the fastest growing economy in the world-has been suffering from increasingly severe air pollution since the 1980s.Related to this, haze occurrence in China, on the whole, has continued to grow during the past several decades, especially after 1980 [1]. Today, haze is a frequent phenomenon in most areas of eastern China, leading to adverse economic as well as human health impacts [2].Broadly, four severe haze regions in China are recognized: Beijing, Tianjin, Hebei province (abbreviated to Jing-Jin-Ji) and its surroundings [3][4][5], Yangtze River Delta, Pearl River Delta, and the Sichuan Basin.As one of the most important urban agglomerations in China, the Jing-Jin-Ji region and its surroundings have attracted considerable attention recently, because of the serious pollution episodes it has experienced since 2013.Multisource observations that can characterize the haze process in Jing-Jin-Ji and its surrounding areas have been used to study the temporal and spatial variation of haze, meteorological conditions, and the chemical components of haze [6][7][8][9][10][11][12].Based on these extensive observational studies, continuous studies of the resultant pollution emissions inventory have also been conducted 2 Advances in Meteorology [13][14][15].In addition, a number of simulation studies using atmospheric models have been carried out to study haze and pollutions processes in China; these studies involve the interactions between meteorological conditions, particle concentrations, and the variation in the transport characteristics of pollutants during the pollution process [16][17][18][19][20].There are two key factors involved in the formation and persistence of haze: one is fine particulate matter (PM 2.5 ) and gas pollutants (O 3 , SO 2 , NO x , etc.) and the other is meteorological conditions.Moreover, when modeling haze, there are uncertainties related to the planetary boundary layer (PBL), which mainly derive from the particular PBL scheme used; and, therein, the PBL height (PBLH), turbulent mixing process, and wind fields are major variables controlling the haze process in the PBL [21][22][23].Therefore, the PBL scheme is a vital impacting factor in terms of modeling the formation and maintenance of haze and air pollution [24,25].A lower PBLH and weaker PBL turbulence diffusion are regarded as key meteorological aspects for haze formation [26].Studies on different PBL parameterization schemes have shown that an accurate depiction of the meteorological conditions within the PBL via an appropriate PBL parameterization scheme is important for air pollution modeling [27][28][29].Some studies have also discussed the importance of the PBL scheme in the modeling of O 3 concentrations, specifically, in the USA and using Weather Research and Forecasting/Chemistry model (WRF-Chem) [30][31][32].These studies also touched upon the possible effects of the PBL scheme on the modeling of PM 2.5 ; however, little is known about whether current PBL schemes are efficient in modeling extremely high PM 2.5 concentrations and haze events over the Chinese mainland.
In order to investigate the abilities of PBL schemes in modeling PM 2.5 over the Jing-Jin-Ji region during serious haze events with high PM 2.5 values and to provide instructive guidance regarding PM 2.5 prediction over this region, separate WRF-Chem model simulations using three popular PBL schemes [Yonsei University (YSU), Mellor-Yamada-Janjic (MYJ), and Bougeault-Lacarrère (Boulac)] were run for haze episodes that occurred in February 2014.After first introducing the methodology, model configuration, and data used, we then evaluate the PM 2.5 simulation results from the three PBL schemes by comparing with observations and analyze the related meteorological fields.Finally, conclusions are drawn regarding the impacts of the PBL on PM 2.5 simulation, along with a discussion on the possible underlying physical mechanisms involved.

Model Introduction and Configuration.
The WRF-Chem model is a fully coupled "online" model, with its air quality component fully consistent with the meteorological component [33,34].Version 3.5 of WRF-Chem was employed in this study.Two nested domains (Figure 1) were used in the simulation with grid spacing of 27 km and 9 km, respectively.The inner domain was centered at 115 ∘ E, 35.5 ∘ N on a Lambert map projection.Considering the regional transmission of PM 2.5 during haze processes, the main research area of domain 2 ranged over 111 ∘ E-120.5 ∘ E, 34.5 ∘ N-42.5 ∘ N, containing the whole Jing-Jin-Ji area and its upstream region including most areas of Shanxi and Shandong provinces and part of Henan province-both regarded as contributors to Jing-Jin-Ji's pollution.The research area is abbreviated as 3JNS hereafter.The two domains used the same 35 vertical levels extending from the surface to 10 hPa, and the layer heights within PBL are shown in Table 3.The simulation period ranged from 00:00 UTC January 28, 2014, to 00:00 UTC March 1, 2014.The simulation outputs from February 1 to 28 were used to obtain the chemical component balance from pollutant emissions.The CBM-Z chemistry mechanism [35] combined with MADE/SORGAM (Modal Aerosol Dynamics Model for Europe and Secondary Organic Aerosol Model) was applied in each domain, and the Fast-J photolysis scheme [36] coupled with hydrometeors, aerosols, and convective parameterizations was chosen.All domains used the RRTM scheme [37] for longwave radiation, the Goddard scheme for shortwave radiation, the Lin (Purdue) microphysics scheme [38], and the New Grell scheme for cumulus parameterization (Table 1).Three PBL schemes-YSU, MYJ, and Boulac-were adopted in the model runs to compare the modeling results of PM 2.5 .

Emissions Instruction.
The anthropogenic emissions of chemical species, with resolution of 0.1 ∘ × 0.1 ∘ , came from the Multiresolution Emissions Inventory for China (MEIC) for 2010 (http://www.meicmodel.org/),which was developed in 2006 for the Intercontinental Chemical Transport Experiment-Phase B (INTEX-B) mission [13].The inventory includes 10 major kinds of pollutants and greenhouse gases and more than 700 kinds of anthropogenic emissions, which can be divided into five sources: transportation, residency, industry, power, and agriculture.According to the INTEX-B inventory, the main pollutants in China that year were SO 2 , NO x , CO, NMVOC, PM 10 , PM 2.5 , BC, and OC.This 2010 emissions inventory has been validated as credible and widely used in studies of pollution in China [14,15,39].

Data Descriptions.
The National Centers for Environmental Prediction (NCEP) reanalysis data (resolution: 1 ∘ × 1 ∘ ) were used for the model's initial and boundary conditions.There are 88 MICAPS stations in 3JNS.
In order to explore the PBL schemes performance in different areas, five stations, Beijing (under the Yan Mountain), Taiyuan (on the west side of Taihang Mountain), Zhangjiakou (in the northwest of 3JNS), Cangzhou (the coastal station), and Xingtai (the east foot of Taihang Mountain) were picked up to represent five different categories of topography and land surface in the 3JNS.The location of their abbreviations is displayed in Figure 1.The topographic basemap of Figure 1 was downloaded from http://www.noaa.gov/.
2.4.Three PBL Schemes' Introduction.PBL schemes can be classified as local or nonlocal closure schemes [40], with the former obtaining the turbulent fluxes of each grid from mean variables and the latter by considering the grid and its surroundings.Additionally, nonlocal schemes are able to simulate the fluxes and profiles of the convective boundary layer.The YSU PBL scheme-an improved version of the Medium-Range Forecast (MRF) scheme, with a critical bulk Richardson number of 0.25 over land-is a revised vertical diffusion package with a nonlocal coefficient in the PBL.Compared with the MRF scheme, it increases boundary layer mixing in the thermally induced free convection regime and decreases it in the mechanically induced forced convection regime.In addition, this scheme is also a relatively mature scheme that is able to simulate a realistic structure of the PBL in the WRF model [41,42].The MYJ PBL scheme is a turbulent kinetic energy (TKE) local closure scheme that defines the eddy diffusion coefficients by forecasting the TKE.This scheme is suitable for all stable and weakly unstable boundary layers [43].The Boulac scheme, regarded as a local closure scheme, has long been regarded as satisfactory in terms of its performance in orography-induced events [44].These three PBL schemes are widely used in mesoscale or weather-scale modeling, and their respective merits or shortcomings have been reported in previous studies.They were also selected for use in the present reported study.

Results and Discussion
3.1.Evaluation of Surface PM 2.5 .To validate the efficiencies of the three PBL schemes in simulating PM 2.5 in the Jing-Jin-Ji region, the spatial distribution of the modeled PM 2.5 values is compared with observations for a severe and long-lasting haze episode in this region.Figure 2 displays the averaged PM 2.5 distribution from 00:00 UTC February 21 to 00:00 UTC February 25, together with the observed values during the same period.The period-averaged PM 2.5 values reached 300-500 g m −3 at observation sites over this region (marked with circles in Figure 2), and the instantaneous values were even higher; the PM 2.5 concentration in some cities (e.g., Beijing, Xingtai, and Tangshan) even reached above 500 g m −3 (Figure 3).Furthermore, as shown in Figures 2 and 3, cities in southern Hebei province endured more severe pollution than northern areas (e.g., Chengde and Zhangjiakou).For this haze period, the model results using the three PBL schemes were all reasonable; the observed and simulated distributions of PM 2.5 showed reasonable consistency.The differences in distributions between the YSU, MYJ, and Boulac schemes were very small.To evaluate the accuracies of the three PBL schemes in modeling the variation in PM 2.5 , 10 representative cities in 3JNS were selected (locations displayed in Figure 2), and their hourly variations in PM 2.5 concentration, as modeled using the three PBL schemes, were compared with observations for the period from 00:00 UTC February 1 to 00:00 UTC March 1 (Figure 3).The results show that all three PBL schemes produced similar representations of the real variation in PM 2.5 for the whole of February, and the differences in modeling values by these three schemes were very little.As the concentration of PM 2.5 is the primary indicators in haze periods, it can be seen from Figure 3 that there were two main haze events in February: one from February 13 to 15 and the other from February 21 to 25.The start and end points of these two events were each modeled well using the three PBL schemes.However, as the simulated conditions of the second event (February 21 to 25) were more accurate, this one was chosen as the research period in this study.In terms of the simulations at individual stations, eastern cities (e.g., Hengshui, Cangzhou, and Chengde) produced better simulation results than western cities (e.g., Zhangjiakou and Baoding) for this event overall, suggesting that the PBL schemes possess properties that are more suited to simulating the PM 2.5 concentration in particular localities.As for how model behaves for particular localities (plains, mountains, or coastal areas, etc.) by using these 3 PBL schemes, we will discuss this below.
Four statistical indicators [mean bias (MB), normalized mean bias (NMB), normalized mean error (NME), and root mean square error (RMSE)] of the haze episode, clean days, and whole month averaged over 3JNS were calculated to evaluate the abilities of the three PBL schemes in simulating PM 2.5 (Table 2).The mean and extreme values of haze and clean periods using each PBL scheme are also displayed in Table 2.The results show that the PM 2.5 modeled during the haze episode was better than that for the whole month.NB and NMB values of less than zero indicate that the model results were an underestimation of the actual situation.The YSU, MYJ, and Boulac schemes underestimated the PM 2.5 concentration during daytime but overestimated it at night, the reason which will be discussed later.On the whole, the Boulac scheme produced the least bias for haze episode compared with the other three schemes, followed by the YSU scheme and MYJ scheme.The MB, NMB, NME, and RMSE values further illustrate that the YSU, MYJ, and Boulac schemes differed little in terms of their simulation of the PM 2.5 concentration during haze.

Relationship between PBL Meteorology and PM
The daily averaged values of PM 2.5 concentration, surface wind speed, PBLH, and vertical diffusivity at level 8 (Table 3) in 3JNS for the whole of February are shown in Figure 4.The PM 2.5 values were determined by averaging the PM 2.5 data of 48 observation stations (Figure 1), and the wind speed values were the average of data of 88 CMA surface monitoring stations in the same area.The three PBL schemes all showed similar trends as those observed.As indicated by the results in Figure 4, the PM 2.5 modeled using the YSU, MYJ, and Boulac schemes also showed very little difference.The three schemes all simulated similar trends for surface wind speed, which were in agreement with the observed trend, though they were all higher than observed.All three schemes showed that the PM 2.5 concentration possessed an accurate inverse relationship with wind speed in terms of the daily averaged variation trend.The daily variation in PM 2.5 concentration also possessed a good inverse relationship with the PBLH (averaged over 48 sites, the same for vertical diffusion), which    suggested that a lower PBLH is an essential prerequisite for haze episodes; but when the PBLH is lower than a certain value, such as 400 m, its relationship with PM 2.5 is not so close.Considering their different diagnoses, the specific values between different PBL schemes are not comparable, so the focus here is the relationships between PM 2.5 and PBL meteorology.The anticorrelation between the daily PM 2.5 and vertical diffusivity of the YSU, MYJ, and Boulac schemes was even weaker than that between PM 2.5 and the PBLH, indicating that the impact of local vertical diffusivity on the time scale of the daily averaged change trend of PM 2.5 is limited.Nevertheless, its impact on the hourly change of PM 2.5 during the daytime is clearer and more important, which will be discussed in Section 3.4.
To illustrate the modeling performance by using three PBL schemes in different topographies, Figure 5     which can represent five topographies in the 3JNS (Figure 1).As to PM 2.5 concentrations and wind speed, PBL schemes can depict appropriate variation trends compared with observation, and they showed a good negative correlation with each other.The difference of PM 2.5 concentration in YSU, MYJ, and Boulac schemes is still little in separate stations; meanwhile, the modeling results in different terrain contain certain differences.In this haze process, the modeling PM 2.5 concentrations in some stations are slightly higher than the observations (Beijing, Taiyuan), while some are lower (Zhangjiakou, Xingtai), and the eastern coastal city (Cangzhou) performed well in this simulation.It is worth mentioning that the Xingtai station, representing the eastern foot of Taihang Mountain, has obviously lower simulating PM 2.5 concentration than the observation by the three schemes, which can be mainly owing to the extremely higher simulation of wind speed (Figure 5).Compared with near stations on the eastern Taihang Mountain, Shijiazhuang and Handan also have similar phenomena (lower simulated PM 2.5 and higher simulated wind speed).It should be noted here that the higher simulated wind speed is one main but probably not the only reason contributing to the higher simulated PM 2.5 .In conclusion, the performance of schemes in the eastern root of Taihang Mountain, the most polluted region by haze in China, was relatively poor due to its specific terrain and complex PBL meteorology.The modeling results in the eastern plain stations (Cangzhou, etc.) of the 3JNS were better than the west (Zhangjiakou, Xingtai, etc.) as mentioned in Section 3.1.
Figure 6 displays the hourly variations of area mean PM 2.5 , wind speed at 10 m, the PBLH, and vertical diffusivity at level 8 (Table 3) of the three PBL schemes during the haze episode (total duration: 120 hours).The YSU scheme simulated the lowest concentration, followed by the Boulac and MYJ schemes.It can be seen from this figure that the PM 2.5 concentrations simulated using the YSU, MYJ, and Boulac schemes all possessed good inverse relationships with wind speed at 10 m, the PBLH, and vertical diffusivity.After sunrise, with the strengthening of solar radiation, the turbulent diffusivity within the PBL continued to improve, the PBLH and wind speed also increased, and all three variables reached their maximum at about 07:00 UTC (local 3 o' clock in the afternoon).After then, all three variables weakened with solar radiation and remained stable at night (after sunset).For the reasons outlined above, the concentration of simulated PM 2.5 during daytime was lower than at night.In summary, in the model, the effects of vertical diffusion on the hourly change trend of PM 2.5 during daytime are much more important compared with the effects on daily averaged PM 2.5 .

Vertical Profiles of PM 2.5 and Meteorology within the PBL.
The structure of PBL vertical meteorology is very important to particle diffusion, vertical and horizontal transportation, and thus the simulation of PM 2.5 .Air sounding observations are only carried out by the CMA at 00:00 UTC (8 o' clock, local time) and 12:00 UTC (nightfall), meaning observational data in terms of the vertical profiles of meteorological parameters during local daytime-local noon (06:00 UTC) especiallyare not available and therefore cannot be used for model validation in China at present.Accordingly, Figure 7 only compares the modeled PM 2.5 concentration, wind speed, and vertical diffusion using the three PBL schemes.Each value of the profile was first averaged over the stations in 3JNS and then averaged over the duration of the haze process (120 hours).The model levels and their corresponding heights are displayed in Table 3.It can be seen from Figure 7 that the differences among the profiles of the YSU, MYJ, and Boulac schemes were small, ranging from 160 g m −3 to 175 g m −3 .The discrepancies in the PM 2.5 concentrations and PBL variables among different PBL schemes were mainly apparent beneath level 11 (height of approximately 402 m).Just under this height local diffusion was the strongest, indicating that local vertical diffusion occurring mainly from 100 m to 400 m and heights below 400 m were important for the PM 2.5 simulation.The results also showed that the surface PM 2.5 concentration was affected by the wind speed and diffusion collectively throughout the whole PBL (especially under 400 m), rather than the surface only.

Diurnal Variation of Surface PM 2.5 and Vertical Diffusion.
The diurnal variations of vertical diffusion and PM 2.5 concentration of haze and clean days using the three PBL schemes are displayed in Figures 8(a)-8(d).These figures show exactly how diffusion affected the PM 2.5 trend during the course of one day, from 00:00 UTC to 23:00 UTC.The values of PM 2.5 and diffusion were both averaged over 48 PM 2.5 stations in the 3JNS, and each hour was also averaged during haze (February 21 to 25) and clean periods (February 3 to 5) separately.Most stations' daily averaged PM 2.5 of observations were above 200 g m −3 in the "haze" and under the 50 g m −3 in the "clean" days.The diurnal variation produced by the three PBL schemes was exactly the same as shown in Figure 6.The PM 2.5 simulated using the YSU, MYJ, and Boulac schemes was obviously lower than observed during daytime, and their diurnal variation of PM 2.5 disagreed with and even contrasted with the observation especially for haze days.It can be concluded that the three PBL schemes might overestimate the vertical diffusion process in the 3JNS region, leading to lower simulated surface PM 2.5 and negative errors during daytime, particularly when severe haze occurred.There are two reasons for this probably strong diffusion and lower PM 2.5 during daytime.The direct radiative feedback of aerosols may lead to weaker diffusion, a more stable atmosphere, and higher surface PM 2.5 when the PM 2.5 concentration is higher than a certain threshold [26,45].However, this feedback was not calculated in the present study.Besides, it was the calculation methods with respect to vertical diffusion by the three PBL schemes that led to stronger particle diffusion and lower surface PM 2.5 than was actually the case in the real atmosphere.
It should be noted that the different representations of vertical diffusion in these PBL schemes might have different impacts on PM 2.5 simulation under different conditions of atmospheric stability in different regions.So here, the same five stations mentioned above were picked up again to illustrate the modeling result of diurnal variations over different topography (Figure 9).For the small difference of vertical diffusivity between haze and clean period at the five stations, the figures of diffusion were ignored here.Though there is no significant pattern in the diurnal variations of observation, this figure also indicated that the simulated diurnal variations of PM 2.5 of specific stations were not as well as daily averaged variations in Figure 5.Despite this, the modeled trends of Taiyuan and the eastern city Cangzhou were better.By the influence of Taihang Mountains, Xingtai simulated lower PM 2.5 in haze days and higher PM 2.5 in clean days compared with observations.Moreover, when the model performed well in haze with high PM 2.5 concentrations (Taiyuan and Cangzhou), it simulated apparently higher PM 2.5 in the clean days with lower PM 2.5 concentrations and vice versa for Zhangjiakou.It seems to be that the little difference of diffusivity calculation between haze and clean days by the PBL schemes calculation might lead to this interesting phenomenon, which is probably the main way to improve PM 2.5 forecasting in complex topography.

Conclusion
To explore the impacts of different PBL schemes on PM 2.5 simulation, three PBL schemes (YSU, MYJ, and Boulac) were applied in the WRF-Chem model to simulate haze episodes   that occurred in the Jing-Jin-Ji region and its surroundings of China.The research area is abbreviated to 3JNS in this paper.The results of the three PBL schemes in simulating the PM 2.5 concentration over 3JNS showed that all these schemes performed similarly with respect to the PM 2.5 trend during a month that included haze episodes.However, among them, the Boulac scheme produced the least bias for haze period, followed by the YSU and MYJ scheme, and these three schemes showed negligible difference in simulating the PM 2.5 concentration.All three PBL schemes simulated similar daily averaged trends in PM 2.5 concentration, which was in agreement with the observation and possessed a good inverse relationship with the PBLH and wind speed-better than with vertical diffusion.All the PBL schemes behave diversely in different terrains.On the whole, eastern plain cities such as Cangzhou and Chengde produced better simulation results than the western cities such as Zhangjiakou and Baoding which are near mountains; the cities under the eastern root of Taihang Mountain produced the worst results in simulating high PM 2.5 ; the modeling results of plain cities were better than the cities under the mountain (e.g., Beijing under the Yan Mountain).The heights under or near the 400 m were found to be very important for PM 2.5 simulation.The effects of vertical diffusion on the hourly change trend of PM 2.5 simulation during daytime were far more important than those on the simulation of daily averaged PM 2.5 .The three PBL schemes might all overestimate the vertical diffusion process in the 3JNS, leading to a lower simulation of surface PM 2.5 and negative errors during daytime-especially when severe haze occurred.In addition, the small gap of diffusivity between haze and clean days by PBL schemes may lead to the errors in simulating PM 2.5 concentrations.It can also be said that the three PBL schemes had not enough ability to distinguish the diffusion between haze and clean days in the complex topography area in China, which may be regarded as an important direction for the improving of PM 2.5 simulation.Since the differences in PM 2.5 concentration among the PBL schemes were small, the exact reasons related to these differences were not discussed in this study.The reasons for the poor reflection of diurnal variation in the PBL schemes, resulting in PM 2.5 errors in numerical models, need to be studied in detail and then adjustments need to be made to improve results for different regions.

Domain 1 2 Figure 1 :
Figure 1: Nested modeling domains (a), the distribution of observation sites within domain 2 ((b) filled circles: PM 2.5 observation sites; open circles: surface meteorological sites; open triangles: upper-air meteorological stations; the dashed-line square area represents the research area (3JNS)), and the topography of the research area (c).

Figure 4 :
Figure 4: Variations of the daily averaged PM 2.5 concentration (a), wind speed near-surface (b), PBLH (c), and vertical diffusivity (d) of the area mean in February.
displays the simulated and observed daily averaged PM 2.5 concentration and wind speed in the whole of February of five stations

Figure 6 :
Figure 6: Hourly variations of the area-averaged PM 2.5 concentration (a), wind speed at 10 m (b), PBLH (c), and vertical diffusivity (d) during the haze process.

Table 1 :
Main physical schemes used in WRF-Chem.

Table 3 :
Model levels and their corresponding heights.