A Hybrid Approach to Forecast Air Quality during High-PM Concentration Pollution Period

In this study, a hybrid approach of combining numerical prediction with statistical analysis was proposed to forecast high-PM10 (aerosol particle with aerodynamic diameter less than 10 μm) concentration events in Beijing, China. This approach was used to forecast the daily PM10 in Beijing from January 1 to December 30, 2013. The WRF-CMAQ modeling system was also applied to simulate Beijing’s PM10 in the same period. The performance of the two methods was then assessed according to the mean bias (MB), normalized mean bias (NMB), normalized mean gross error (NME), mean normalized bias (MNB), mean normalized gross error (MNE), and root mean square error (RMSE). The results demonstrate that both methods perform well during low-PM10 concentration periods (PM10 concentration < 250 μg/m), the MB, NMB, NME, MNB, MNE and RMSE for hybrid approach during low-PM10 concentration periods were 26.15, 24.88%, 41.94%, 43.23%, 56.35% and 61.67, respectively. The MB, NMB, NME, MNB, MNE and RMSE for CMAQ during low-PM10 concentration periods were –6.04, 57.47%, 41.49%, 21.52%, 55.64% and 60.11, respectively. While the MB, NMB, NME, MNB, MNE and RMSE for CMAQ during high-PM10 concentration periods (PM10 concentration ≥ 250 μg/m) were –162.87, –50.37%, 50.37%, –49.86%, 49.86% and 175.93, respectively. The MB, NMB, NME, MNB, MNE and RMSE for hybrid approach during high-PM10 concentration periods were –30.3, –9.37%, 23.21%, –8.21%, 24.25% and 97.37, respectively. The hybrid approach shows significant improvement in accuracy during high-PM10 concentration periods.


INTRODUCTION
In recent years, in China, pollution from high concentrations of particulate matter (PM) (C PM10 ≥ 250 μg/m 3 ) has been receiving growing attention.It not only affects the development of cities but also produces a serious public health threat (Cheng et al., 2007).Air pollution control is needed to prevent the situation from rapidly becoming worse.The Chinese government has committed to reducing pollution from particulate matter in various ways, but pollution is still not alleviated.In Beijing, for example, the average annual concentration of PM 2.5 (aerosol particle with aerodynamic diameter less than 2.5 μm) was 89.5 μg/m 3 in 2013 (BMEPB, 2014), which is much higher than the World Health Organization standard (35 μg/m 3 ) and America National Ambient Air Quality standard (15 μg/m 3 ).Chinese people suffer greatly from particulate matter air pollution.Thus, just like weather forecast, air pollution forecast is necessary to allow people to take preventative measures against exposure during high-pollution events (Kolehmainen et al., 2001).Additionally, accurate air quality forecast can offer tremendous societal and economic benefits by enabling advanced planning for individuals, organizations, and communities to reduce pollutant emissions and their adverse health impacts (Zhang et al., 2012).
The current widely used forecast methods are statistical and numerical models, which determine the relationship between PM concentrations and certain input parameters (Li et al., 2011).Numerical model forecast is based on atmospheric physics and atmospheric dynamics theory.This method establishes a numerical model, which simulates pollutant transport and diffusion, and it predicts the dynamic spatial and temporal distribution of air pollutants through a large number of numerical calculations.The numerical forecast method is currently the most popular air quality forecast system in most countries.The United States (Lee et al., 2007) has used Models-3/CMAQ, the third generation air quality modeling system developed by the United States Environmental Protection Agency, to forecast O 3 and PM 2.5 average concentrations and maximum hourly concentrations of O 3 in a subsequent 24-h period.Japan (Chatani et al., 2011) used the CMAQ framework to simulate O 3 , NO x , and air quality concentrations of particulate matter.The Netherlands (Manders et al., 2009) investigated the ability of the chemical transport model LOTOS-EUROS to forecast PM 10 concentrations.The statistical forecast method is based on the relationship between a set of meteorological predictors and the values of the concentrations at different monitoring stations (Vautard et al., 2001).Statistical forecasts use a variety of mathematical models such as linear or nonlinear regression (Cobourn and Hubbard, 1999;Cobourn, 2007), artificial neural network (Abdul-Wahab et al., 2002;Nejadkoorki et al., 2012;Antanasijevic et al., 2013), and fuzzy logic (Shad et al., 2009).Yetilmezsoy et al. (2012) proposed a prognostic approach, which is based on a fuzzylogic model, to estimate suspended dust concentrations related to PM 10 in a specific residential area in Kuwait with high traffic and industrial influences.Vlachogianni et al. (2011) developed forecast models based on stepwise multiple linear regression (MLR) for Athens and Helsinki to forecast the maximum hourly concentrations of PM 10 and NO x , and the daily average PM 10 concentration of the next day.Cobourn (2010) developed an enhanced PM 2.5 air quality forecast model based on nonlinear regression (NLR) and backtrajectory concentrations for use in the Louisville, Kentucky metropolitan area.The two methods mentioned above are widely used and relatively accurate.
In China, many researchers apply numerical models to predict the concentrations of pollutants.Xu et al. (2005) proposed a new way of combining dynamics and statistics by establishing a statistically correcting model (CMAQ-MOS) to forecast regional air quality.This utilized the relationship of CMAQ outputs with corresponding observations; they tested the forecast capabilities as well.Wang et al. (2012) applied an urban air quality forecast system to predict air quality in Shanghai and Nanjing; their model was based on the latest weather research forecast and chemistry model (WRF-Chem).A regional haze weather forecast system, based on the Regional Atmospheric Environment Modeling System (RegAEMS), was applied for regional haze weather forecast in the Yangtze River Delta.These studies indicated that the numerical forecast method performs well during low-PM concentration periods (PM 10 concentration < 250 μg/m 3 ); however, during high-PM concentration periods (PM 10 concentration ≥ 250 μg/m 3 ) forecast deviation is relatively large, and the accuracy of the numerical model is far from the desired results.There are a few explanations for the deviation.First, both temporal and spatial resolution of China's current pollutants emission inventory, which is needed in the numerical model, is much lower, and the species of pollutants are so inadequate that it cannot provide enough information for the numerical models to achieve the accurate results.The uncertainties in emission estimates for China vary considerably across sectors and source types, such that large uncertainties may still remain in individual locations (Zhang et al., 2009;Zheng et al., 2009aZheng et al., , b, 2011)).Second, there is uncertainty regarding which modeling mechanism is applicable to China, as air pollution in Chinese cities is such a serious problem due to very high concentrations of a variety of air pollutants, and the chemical processes among them are extremely complicated (Fang et al., 2005;Geng et al., 2007;Zhao et al., 2006).OH concentrations are significantly higher compared with the global average (Ren et al., 2004;Ren et al., 2002;Shao et al., 2004).These factors lead to atmospheric conditions in China that are much different from countries in North America and Europe.Numerical models are currently more widely used, but they are generally designed for specific locations and provide limited output (Manders et al., 2009).Although there are examples that use a statistical forecast model to predict concentration of pollutants directly (Yi et al., 1996;Wang et al., 2003), the number is very limited due to the lack of accuracy.A widely accepted explanation for the inaccuracy is that the statistical analysis forecast in these studies does not incorporate the meteorological data of the forecast day, which is vital information.Because air quality is greatly affected by weather conditions, during days when weather conditions are unfavorable for the transport and diffusion of pollutants, air quality can be very poor (Liu et al., 2007).This may also explain why the air quality forecasts did not match the observed data.In fact, a forecast model has yet to be found that is sufficiently accurate to be used routinely by air quality managers in China.
On the other hand, although the Chinese Government has taken various measures to reduce particulate emissions, air quality has not been radically improved.In January 2013 several high-PM concentration events occurred.Pollutant concentrations reached an unprecedented value, and, since then, the demand for forecast accuracy of high-PM concentration pollution has intensified.As a result, it is of great importance to develop an accurate, high-PM concentration forecast system that is suited for China to minimize negative effects from air pollution (Cheng et al., 2012).
To effectively protect human health, the real-time air quality status is needed, especially for high-PM concentration periods.We present in this article, a hybrid approach; it applies numerical simulation combined with statistical analysis to establish a high-PM concentration pollution forecast system.The numerical model we utilize is the weather research and forecast model (WRF).From this, we can obtain meteorological data for the forecast day, such as 24-h wind speed, 24-h temperature, and 24-h pressure; statistical analysis forecast used alone was unable to provide these data and therefore was incomplete.In this study, we considered simulated and actual local conditions, such as regional pollution characteristics, meteorological and ground conditions, variations in pollution concentrations, and the variability of source emissions in China, to conduct the forecast work.Beijing was chosen as a case study, and the hybrid forecast system was used to predict PM 10 concentration from January 1 to December 30, 2013.WRF/CMAQ was also used to simulate PM 10 concentration from January 1 to December 30, 2013.The performance of the two methods was assessed according to the mean bias (MB), normalized mean bias (NMB), normalized mean gross error (NME), mean normalized bias (MNB), mean normalized gross error (MNE), and root mean square error (RMSE).

METHODOLOGY
To solve the problem of low forecast accuracy of high-PM concentration pollution, this paper focuses on the study of air quality forecast during periods of high-PM concentration pollution.The forecast system was integrated with various technical methods based on the characteristics of high-PM 10 concentration pollution, transport of pollutants, the relationship between particulate matter chemicals, and weather conditions, among other findings.The forecast system serves as an automatic diagnostic of high-PM 10 concentration pollution and as an air quality quantitative forecast method.

Database
The air quality data used to build and test the forecast system was obtained from the Chinese Ministry of Environmental Protection and the Beijing Municipal Environmental Monitoring Center data center.Historical meteorological data were collected from the China Meteorological Administration.The air quality data used to build and test the forecast system was obtained from the Chinese Ministry of Environmental Protection and the Beijing Municipal Environmental Monitoring Center data center.Historical meteorological data were collected from the Beijing meteorological station.8087 effective PM 10 datasets collected from 12 monitoring stations (AoTi, Changping, Dingling, DongSi, GuCheng, Guanyuan, Huairou, NongZhanGuan, Shunyi, Wanliu, Wanshouxigong and TianTan) in Beijing from 2000 to 2011 were used as air quality input.2025 corresponding meteorological datasets were collected from Beijing meteorological station (ID54511).Table 1 shows the detailed information of these datasets.

Meteorological Model: WRF
In this paper, WRF (Weather Research and Forecast) was used to forecast meteorological elements.The input parameters of the meteorological model are as follows: in the vertical direction, the terrain-following σ-coordinate system is applied, vertical stratification is 27 layers, and numerical integration is 60 s.The physical process uses Lin et al. (1983) microphysics, Kain-Fritisch (new Eta) integration scheme, and Dudhia cloud radiation scheme and soil stratification is four layers.The meteorological data employed is predicted from the U.S. National Centers for Environmental Prediction global forecast system (GFS).Spatial resolution is 1° × 1° and the time interval is 3 h.For this forecast system, the simulated region of meteorological model is shown as Fig. 1.As shown in Fig. 1, three nesting levels of 36 km, 12 km, and 4 km spatial resolution on a Lambert map projection have been chosen for the meteorological simulation.The 36-km spatial resolution covers most of the surrounding provinces under investigation; the domain with 12-km spatial resolution covers the North China Plain; the inner sub-domain with 4km spatial resolution covers the whole Beijing area.

Weather Type Classification
The weather type-classification method includes SSC (spatial synoptic classification) and TSI (temporal synoptic index) (KALKSTEIN et al., 1996).This paper uses the TSI method, comprising the principal component analysis (PCA) and the average linkage clustering model (Chad et al., 2007).Principal component analysis decomposes the total variance of original variables X 1 , X 2 , ..., X p into p uncorrelated additive variance of variables Y 1 , Y 2 , ..., Y p , that is The contribution rate of components, k, is: The cumulative contribution rate of the first m principal components is: It shows that X1, X2, ..., Xp are capable of providing comprehensive information of the first main m components Y1, Y2, ..., Ym.Six meteorological elements (surface temperature, dew point, sea level pressure, total cloud cover, wind direction, and wind speed) are selected to represent instantaneous weather characteristics (Chad et al., 2002).According to meteorological variables obtained from WRF model, we select TSK (land surface temperature), POTEVP (Accumulated Potential Evaporation), PSFC (Surface Skin Pressure), RAINC (Accumulated Total Cumulus Precipitation), and U(X-wind Component, V(Ywind Component), W (Z-wind Component) as the components representing the six meteorological elements mentioned above (Zheng et al., 2004).The selected components and their loads are shown in Table 2   The cumulative variance contribution rate of the former four main components reached 88.32%; the score matrix for these components acts as the input data for the average link clustering model.The average link method defines square of the distance between two types of classes as the average square of the distance between these two classes in pairs.That is: where d ij is the Euclidean distance between the sample x i and the sample x j .
In this paper, the number of clusters was tested; the cluster results are divided into 12 categories (category equations can be more (three classes)), and the data within each category were more evenly distributed.The proportion of the number of days within each category are shown in Table 3.
Individual categories contain less data after clustering, such as TYPE 3, TYPE 5, and TYPE 6, and because the

High-PM 10 Concentration Pollution Weather Qualitative Judgment
To forecast different levels of pollution more accurately, especially for the values of C PM10 in heavily-polluted conditions, we established forecast equations for each level of pollution using the logistic regression principle.We predetermine air quality qualitatively according to meteorological factors of forecast samples.In this study, atmospheric particulate matter pollution levels are divided into three grades: level 1 C PM10 ∈ [0,250 μg/m 3 ), level 2 C PM10 ∈ [250 μg/m 3 , 370 μg/m 3 ), and level 3 C PM10 ∈ [370 μg/m 3 , +∞).The steps for identifying the PM 10 pollution levels are as follows: (1) if the C PM10 value of the forecast day is less than 250 μg/m 3 , then the pollution level is grade 1; (2) if C PM10 is more than 250 μg/m 3 but less than 370 μg/m 3 , then the pollution level is grade 2; (3) if C PM10 is more than 370 μg/m 3 , then the pollution level is grade 3. P denotes the probability that an event will happen.The basic formula for the logic conversion is: with logit (0.5) = 0 as the center of symmetry, as shown in Table 4. Identification process is shown in Fig. 2.

High-PM 10 Concentration Pollution Weather Quantitative Forecast
In this paper, land surface temperature, accumulated potential evaporation, land surface pressure, and other meteorological elements (shown in Eq. ( 4) shows the stepwise regression, where C PM10 j is the PM 10 concentration with units μg/m 3 in the j-th time, a is the constant term in the regression equation, b i is the coefficients term of the independent variables in the i-th time in regression equation, and x ij is the i-th meteorological element on the j-th day.

Model Integration
The high-PM concentration quantitative forecast system integrates the weather type-classification model, high-PM  concentration pollution weather qualitative judgment model, and high-PM concentration pollution weather quantitative forecast model; it uses the Matlab platform to perform secondary linear factor calculations and enters historical meteorological elements and C PM10 automatically.For the actual forecast input data, we chose the time range starting at 20:00 on the first day and ending at 20:00 the following day, the meteorological elements from WRF, and the C PM10 of the five consecutive hours prior to the forecast time range.We apply the quantitative forecast system to predict C PM10 of the next 24 h.The forecast process is shown in Fig. 3.

RESULTS AND DISCUSSION
The development of the forecast system described above was applied to the research area of Beijing to establish a high-PM 10 concentration pollution-forecast system for Beijing.This hybrid forecast system was used to forecast dozens of historical high-concentration pollution events and the real-time PM 10 concentrations from January 1 to December 30, 2013.WRF/CMAQ was also used to conduct PM 10 concentration forecast from January 1 to December 30, 2013 for comparison.

Historical Pollution Process Simulate Forecast Results
To inspect the high-PM concentration pollution prediction effect of this technical approach, we randomly selected more than 2,800 sets of pollution data from 2000 to 2011.These datasets were used to conduct simulated forecast (including a dozen periods of high-PM concentration pollution).Among those historical datasets, high-PM concentration pollution occurred frequently from October to November 2002.From October 15 to November 5, there were seven days during which daily average PM 10 concentrations were higher than 250 μg/m 3 and two days during which daily average PM 10 concentrations were higher than 350 μg/m 3 .Fig. 4 shows the daily average forecast results.

Real-Time Forecast Results of this Forecast System
To evaluate the performance of this forecast system, we compared the forecast PM 10 concentrations with the groundbased observations in Beijing.For good evaluation of this forecast system, the same 12 air quality-monitoring stations located within the urban area and suburbs areas of Beijing were selected, including AoTi, Changping, Dingling, DongSi, GuCheng, Guanyuan, Huairou, NongZhanGuan, Shunyi, Wanliu, Wanshouxigong and TianTan, corresponding with the quality-monitoring stations which were selected to establish forecast system.The observed hourly PM 10 results from these 12 monitoring stations were averaged and then compared with the predicted daily PM 10 concentration in Beijing.Fig. 5 shows the comparison of observed data and forecast data daily from January 1 to December 30, 2013.
Through comparison of the forecast results with realtime daily average PM 10 concentrations, we can see that the performance of this forecast system is good; it reflects the increasing and decreasing trends of the PM 10 concentration very well.Overall, forecast data are slightly higher than the observed data, possibly because the forecast system focuses on high concentrations of particulate contamination; whereas, the data used to establish the forecast system is mainly from a historical high-PM concentration period.Its performance in high-PM concentration periods is better than in low-PM concentration periods.
During the end of March to early April, forecast values were slightly lower than the observed values.The forecast  system did not take into account the impact of sandstorm events, which increased the airborne particulate concentrations at that time.

Simulated Results of CMAQ
We applied CMAQ to simulated real-time daily average PM 10 concentrations from January 1 to December 30, 2013.The emission inventory used in the study is the MEIC inventory developed by Tsinghua University.It contains monthly emissions for China and is available on a 0.25° resolution for 2008 and 2010 through http://www.meicmodel.org (Mijling et al., 2013).Meteorological data are from WRF and the parameter settings of model simulation is consistent with the WRF parameter settings.Fig. 6 shows the comparison of observed data and forecast data of CMAQ.
From the forecast results of CMAQ, we can see that in low-PM concentration periods, the model performs well; it can reflect the increasing and decreasing trend of PM 10 concentration, whereas in high-PM concentration periods, forecast values are much lower than observed value.Therefore, CMAQ is unable to accurately forecast high-PM concentrations.

Comparison of Two Forecast Results
According to the U.S. EPA model evaluation protocol, the mean bias (MB), normalized mean bias (NMB), normalized mean gross error (NME), mean normalized bias (MNB), mean normalized gross error (MNE) and root mean square error (RMSE) are used to perform statistical analysis of the two forecast results.Table 8 shows the comparison of the two forecast results from January 1 to December 30, 2013.Table 9 shows the comparison of two forecast results during low-PM concentration periods (C PM10 < 250 μg/m 3 ) and Table 10 shows the comparison during high-PM concentration periods (C PM10 ≥ 250 μg/m 3 ).
where Model -Forecast data; Obs -Observed data; Nnumber of data pairs The MB and NMB provide an indication of whether the predictions are over or under estimated, while RMSE provide a good indication of how close the modelled and observed values are.A negative MB and NMB value indicate underestimation, whereas a positive MB and NMB indicate an overestimation of the predicted PM 10 concentrations.Higher RMSE values indicate higher error, which shows poorer agreement of the forecast and observed values (Sayegh et al., 2014).The results of this hybrid forecast system are overpredicted with the average NMB of 16.05% and the NME of 37.55% in the whole year, respectively.While the results of CMAQ are underpredicted with the average NMB of -17.75%.In low-PM 10 concentration periods, both the hybrid forecast system and CMAQ overpredicted with the average NMB of 24.88% and 54.47%, respectively.Though difference exists in NMB, NME and RMSE in the whole year, CMAQ's performance is slightly better than the hybrid forecast system with respect to RMSE during low-PM 10 concentration periods.
During high-PM 10 concentration periods, the MB, NMB, NME, MNB, MNE and RMSE for CMAQ were -162.87,-50.37%, 50.37%, -49.86%, 49.86% and 175.93, respectively.While the MB, NMB, NME, MNB, MNE and RMSE for this hybrid forecast system were -30.3, -9.37%, 23.21%, -8.21%, 24.25% and 97.37, respectively.A statistical From the comparison, we can see that both forecast methods produce very similar results.In low-PM 10 concentration periods, CMAQ's performance is slightly better than the hybrid forecast system, whereas during high-PM 10 concentration periods, performance of the hybrid system is better.It is important to accurately forecast high-PM 10 concentrations to ensure that people can take protective measures in a timely fashion; therefore, during low-PM 10 concentration periods, forecast data could be based on CMAQ results and in high-PM concentration pollution periods, forecast could be based on the results of the hybrid forecast system.

CONCLUSIONS AND RECOMMENDATIONS
In this study, we established a hybrid forecast system to forecast periods of high-PM concentrations.Using a numerical model and a variety of statistical methods, this hybrid forecast system predicts high-PM pollution events more accurately than current forecast methods.It combines the strengths of various methods while avoiding the disadvantages found when statistical forecast methods are used alone.The weather type-identification and classification models based on meteorological clustering were added before quantitative forecast of pollution; this addressed the lack of physical representation in the statistical forecast methods.Quantitative forecast models based on different weather types and different pollution levels made forecast results during high-PM concentration periods more accurate.Compared with the forecast result of CMAQ, performance of this hybrid forecast system was better during high-PM concentration pollution periods with respect to MB, NMB, NME, MNB, MNE and RMSE values.During high-PM 10 concentration periods, the MB, NMB, NME, MNB,50.37%,49.86% and 175.93, respectively.While the MB, NMB, NME, MNB, MNE and RMSE for this hybrid forecast system were -30.3, -9.37%, 23.21%, -8.21%, 24.25% and 97.37, respectively.A statistical analysis of the results obtained from the application of the two methods demonstrated that much lower values of MB, NMB, NME, MNB, MNE and RMSE resulted from this hybrid forecast system in high-PM 10 period and lower values indicate lower error.This hybrid forecast system has been doing much better than CMAQ in high-PM 10 concentration periods.During low-PM 10 concentration periods, performance of CMAQ is slightly better than the hybrid forecast system.It is indicated that in low-PM 10 concentration periods, forecast data could be based on CMAQ results and in high-PM concentration period, forecast data could be based on the results of the hybrid forecast system.
However, disadvantage exists in every method.We summarize the disadvantage of this hybrid forecast system as following: (1) The physical and chemical processes of atmosphere were not explicitly incorporated in the hybrid model.This is the congenital defect of statistical methods.However, we have exhausted our ability to choose the physically and chemically meaningful parameters in the processes of building the hybrid model.(2) This method is area specific, not universal.It means that if we apply this model in a new place, we have to spend extra effort to retrain the model by the use of same amount of local data as we did initially, which is very time-consuming.(3) It is data demanding.The building up of this hybrid model requires huge amount of data (usually years of hourly-specific data).This is also the reason why we chose PM 10 instead of PM 2.5 to build this hybrid forecast system.PM 2.5 began to attract attention in China in recent years and the Chinese government only started collecting PM 2.5 concentration data in 2013, there are not enough datasets to establish a forecast system of PM 2.5 with the method in this study.The forecast system this study established is only for the prediction of PM 10 for now.However, with the accumulation of observations in the future, next study will concentrate on the establishment of a forecast system of PM 2.5 as soon as enough data are available.

Fig. 3 .
Fig. 3. Forecast process of the high-PM concentration pollution quantitative forecast system.

Table 1 .
. There Meteorological and air quality data used in the data regression.

Table 2 .
Scores of the Former Four Principle Components Selected.

Table 3 .
Cluster Result of Hourly Atmospheric Figure of Beijing in 2000-2011.
Table 1) were selected as input data for the statistical equation in the quantitative forecast model.Based on the Matlab platform, a stepwise regression forecast equation was established for each different weather type and each different pollution level to calculate the concentration of pollutants quantitatively.The forecast equation based on stepwise regression method is shown in Eq. (4).Coefficient term b i and the constant term correspond to the formula as shown in Table 5 (TYPE 1), Table 6 (TYPE 2), and Table 7 (TYPE 4).

Table 4 .
P Value and Logit (p) Value when Using Logistic Regression.
Fig. 2. Flow Chart of Logistic Regression Procedure.

Table 5 .
Equation Coefficients of the Stepwise Regression in TYPE 1.

Table 6 .
Equation Coefficients of the Stepwise Regression in TYPE 2.

Table 7 .
Equation Coefficients of the Stepwise Regression in TYPE 4. the meteorological elements that were removed automatically from the regression equation.Some values were expressed in scientific notation due to limited column space.

Table 8 .
Comparison of two forecast results from January 1 to December 30, 2013.

Table 10 .
Comparison of the two forecast results during high-PM 10 concentration periods (C PM10 ≥ 250 μg/m 3 ).results obtained from the application of the two methods demonstrated that much lower values of MB, NMB, NME, MNB, MNE and RMSE resulted from this hybrid forecast system in high-PM 10 period and lower values indicate lower error.This hybrid forecast system has been doing much better than CMAQ in high-PM 10 concentration periods.